Intelligent purification process control system for biopharmaceuticals based on big data

Abstract

The biopharmaceutical intelligent purification process control system based on big data includes a data acquisition module, a quality control module, a system management module, and a data visualization and reporting module: wherein the data acquisition module uses sensing technology to collect parameters of the biopharmaceutical purification technology and records them in the control system, the quality control module is used to control and optimize the biopharmaceutical purification process, the system management module is used to maintain the usage of the control system users, and the data visualization and reporting module is used to present the control process and form a report on the control results; the present invention will further add an intelligent control function based on the quality control module, establish a functional relationship between the parameters in the control process and the drug purity through historical data, and achieve the purpose of optimizing production parameters to improve the drug purity; this paper mainly uses CNN network and LSTM network to optimize the parameters in the process, which significantly improves the purity of the drug.

Description

[0001] The invention relates to the field of biopharmaceutical purification, and specifically to a biopharmaceutical intelligent purification process control system based on big data. Background Art

[0002] In recent years, the biopharmaceutical industry has developed rapidly, and the requirements for purification processes have become increasingly higher. Traditional purification processes mainly rely on the experience of operators and some simple control methods. This method seems to be unable to cope with complex production environments and changing market demands. With the rapid development of big data and artificial intelligence technology, it has become feasible and efficient to use machine learning technology to control biopharmaceutical purification processes. Through in-depth analysis of a large amount of historical data and real-time monitoring data, machine learning algorithms can accurately predict the yield, purity and quality of biological products, thereby optimizing parameter settings in the production process and improving overall production efficiency and product quality. The system can achieve precise control of key parameters in the production process through intelligent adjustment and optimization. Specifically, the machine learning algorithm can capture various key factors that affect the purity and yield of drugs, establish a relationship model between parameters and purification results, and monitor and adjust these parameters in real time during the production process. This can not only ensure stable and efficient operation under different production conditions, but also greatly improve the purity and production efficiency of drugs to meet increasingly stringent quality requirements. In addition, the system can continuously optimize the algorithm model by continuously learning and updating data to adapt to changes in production processes. This dynamic adjustment capability enables the system to maintain a high level of stability and reliability when facing production tasks of different batches and conditions. Through the application of big data and artificial intelligence technology, the purification process of the biopharmaceutical industry is moving towards a more intelligent and refined future.

[0003] A biopharmaceutical intelligent purification process control system based on big data includes a data acquisition module, a quality control module, a system management module, and a data visualization and reporting module. The quality control module proposes a technology to improve the purification effect of the system through the CNN-LSTM algorithm. Convolutional neural network (CNN) has a strong feature extraction capability and can effectively identify local patterns and structures in data. In the biopharmaceutical purification process, changes in key parameters (such as temperature, pressure, flow rate, pH value, dissolved oxygen content, etc.) often contain important feature information. Through CNN, useful features can be extracted from these data to provide high-quality input for subsequent predictions. Long short-term memory network (LSTM) is good at processing time series data and can capture long-term dependencies in data. In the biopharmaceutical purification process, parameter changes in the production process are often time-series, and past parameter changes will affect future production results. Through LSTM, the changing trend of these time series data can be effectively predicted, and the intelligent control and optimization capabilities of the production process can be improved. Combining the feature extraction capability of CNN with the time series prediction capability of LSTM can give full play to the advantages of both. In practical applications, CNN is first used to extract features from production data, and then these features are modeled and predicted for time series through LSTM. This can more accurately capture the changing rules of key parameters and improve the prediction accuracy and control effect of the production process. Through the CNN-LSTM network, the system can automatically identify and adjust key parameters to optimize production conditions and ensure that the drug purity and production efficiency in the production process are optimal. In addition, CNN can process a large amount of complex data, extract key features, reduce the dimension of data, and improve data processing efficiency. LSTM can capture long-term dependencies, provide more accurate time series predictions, and help optimize production parameter settings. The system can monitor the key parameters in the production process in real time, make intelligent adjustments based on the prediction results, and ensure the stability and efficiency of the production process. Based on the machine learning model, the system can be optimized and adjusted under different production conditions to improve drug purity and production efficiency and meet increasingly stringent quality requirements. By introducing the CNN-LSTM network, this system can better identify and predict the changes in key parameters in the production process and realize intelligent control and optimization of the biopharmaceutical purification process. This not only improves production efficiency and product quality, but also enhances the stability and adaptability of the system, meeting the needs of the modern biopharmaceutical industry for efficient, safe and high-quality production. Summary of the invention

[0004] In view of the above problems, the present invention aims to provide a biopharmaceutical intelligent purification process control system based on big data.

[0005] The purpose of the present invention is achieved through the following technical solutions:

[0006] A biopharmaceutical intelligent purification process control system based on big data, including a data acquisition module, a quality control module, a system management module, and a data visualization and reporting module, wherein:

[0007] The data acquisition module is used to use sensor technology to collect parameters in the biopharmaceutical purification process and record them in the control system; the quality control module has two functions. First, it groups historical data. Second, through the historical grouped data, it uses machine learning algorithms to establish a functional relationship between the parameters in the control process and the drug purity, thereby realizing intelligent adjustment and optimization of the production process and improving drug purity; the system management module includes system maintenance and monitoring functions and security and authority control functions. The system maintenance and monitoring function plays a role in monitoring system performance and fault detection, and the security and authority control function plays a role in user identity authentication and authority control; the data visualization and reporting module provides an intuitive data visualization interface, displays the data and analysis results in the control process, and generates analysis reports.

[0008] Furthermore, the data acquisition module includes a sensor technology application unit, a real-time data acquisition unit, a data recording and storage unit, a data preprocessing unit, and an interface and communication unit, wherein: the sensor technology application unit is used to monitor and collect temperature, pressure, flow rate, pH value, and dissolved oxygen content in the biopharmaceutical purification process, and the sensor is arranged inside the purification equipment; the real-time data acquisition unit is connected to the data acquisition module through a high-speed data interface to achieve a millisecond-level data acquisition frequency, and transmits the data to the central control system in real time through the data acquisition system; the data recording and storage unit has data storage capabilities, classifies and stores the collected data, adopts redundant storage technology, and regularly backs up the data; the data preprocessing unit preprocesses the data during the data acquisition process, including noise removal, outlier detection and data correction, and marking timestamps; the interface and communication unit supports docking with the quality control module, system management module, data visualization and reporting module, and connects to the remote server through a network interface to remotely monitor and analyze the data.

[0009] Furthermore, the quality control module includes a historical data grouping unit, a machine learning algorithm unit, a model training and updating unit, a data feedback and early warning unit, and a user-defined parameter setting unit, wherein: the historical data grouping unit is used to group historical data according to temperature, pressure, flow rate, pH value, and dissolved oxygen content characteristics (the dissolved oxygen content is the content of oxygen dissolved in water, in mg / L) to facilitate analysis and comparison of production results and provide a data basis; the machine learning algorithm unit analyzes the historical grouped data and uses the machine learning algorithm to establish a functional relationship between the characteristics in the control process and the drug purity, so as to capture the impact of changes in data characteristics on drug purity; the intelligent adjustment and optimization unit performs real-time monitoring and intelligent adjustment of the production process based on the model established by the machine learning algorithm. , the intelligent adjustment and optimization unit can automatically identify and adjust parameters; the model training and update unit can use new data to continuously improve and optimize the model, adapt to changes in the production process, and control the temperature, pressure, flow rate, pH value, and dissolved oxygen content characteristics within the preset range to accurately predict the yield and purity of biological drugs; the data feedback and early warning unit, through the monitoring and analysis of real-time data, when it is found that the parameters in the production process are abnormal and deviate from the preset range, the system can promptly issue an early warning and feedback to the operator to ensure that measures can be taken quickly to avoid production accidents and quality problems; the user-defined parameter setting unit allows the user to adjust the parameter settings according to needs, and through the optimization of machine learning algorithms, combined with historical data and real-time monitoring results, provide personalized parameter optimization solutions to improve production flexibility and efficiency.

[0010] Furthermore, the historical data grouping unit in the quality control module obtains the data of the data acquisition module in the system from the database, and the module reads the historical data set, and the historical data independent variable set is recorded as X = {x1, x2, x3, ..., x p ,…,x n}, the historical data dependent variable set is recorded as Y = {y1, y2, y3, ..., y p ,…,y n}, where x p ,y p Design a feature type set to purify the type. The numbers 1, 2, 3, ..., p, ..., n in the X and Y sets are the 1st, 2nd, 3rd, ..., p, ..., nth data in the historical data, denoted by x p ={xw1,xw2,xw3,xw4,x w5}, where xw1 is the temperature characteristic, xw2 is the pressure characteristic, xw3 is the flow rate characteristic, xw4 is the pH value, and xw5 is the dissolved oxygen content characteristic. p The features are regarded as a set of dependent variables, denoted as y p={yw1,yw2}, where yw1 is the purity of the drug and yw2 is the drug yield, recorded as the classification interval set Among them, 1, 2, ..., n1 are the classification interval numbers, m is the grouping number, and k is the classification method. For categorical data, In this system, three types of classification are used: empirical formula grouping, cluster grouping, and classification tree grouping. The empirical formula grouping method is specifically as follows: determine the specific range of temperature, pressure, flow rate, pH value, and dissolved oxygen content, and set the range of temperature (T) to T min To T max , where T min is the lowest temperature of the purification process, T max The maximum temperature of the purification process is set to P. min To P max , where P min is the minimum pressure of the purification process, P max For the highest pressure of the purification process, set the flow rate (F) in the range of F min To F max , where F min is the minimum flow rate of the purification process, F max For the highest flow rate during purification, set the pH range to pH min To pH max , where pH min The lowest pH value during the purification process, pH max For the highest pH value of the purification process, set the dissolved oxygen content (DO) range to DO min To DO max , where DO min The minimum dissolved oxygen content in the purification process, DO max The highest dissolved oxygen content in the purification process is obtained by establishing an empirical formula S for purity prediction through the heuristic method of polynomial fitting. The data are grouped from small to large according to S, and a total of 20 groups are divided, and the group set is recorded as G m1 , where m is the group number, the maximum value is 20, and the prediction formula is as follows:

[0011] S=aT·bP·cF+d(pH)+e(DO)+r

[0012] Among them, a, b, c, d, e, and r are the parameters fitted by the heuristic algorithm. The clustering method is specifically to transform each element x in the X set i , calculate x i With the group center μ m The Euclidean distance of the center of x i Assigned to the closest group, denoted by Gm2 is the set of the mth group, μ m is the center point of the m group set, μ ml is the center point of any grouping set:

[0013]

[0014] Among them, for G m2 , calculate the updated group center μ m ':

[0015]

[0016] in, G m2 The number of elements in the set is recorded. The classification tree method is specifically a gradient boosting decision tree algorithm. For each feature (temperature, pressure, flow rate, pH value, dissolved oxygen content) data, there is a classification tree model:

[0017]

[0018] Where F0(x) is the initial prediction value, γ is the leaf node prediction value, Le is the loss function, and F X is the predicted value in the X set, r nj is the model prediction residual, is the predicted value of F in the previous step,

[0019]

[0020] And fit a new decision tree, For the input x j Decision tree model:

[0021]

[0022] h X is the new decision tree obtained by fitting the model, updating the model,

[0023] F n (x) = F n-1 (x)+0.1h X (x)

[0024] F n (x) is the latest model prediction value, F n-1 (x) is F n (x) The predicted value of the previous model, the final classification result will be the set of classification methods C = {G mk,m∈{1,2,3,…o},k∈{1,2,3}}, where 1, 2, and 3 in k represent the empirical formula method, clustering method, and classification tree method respectively, and o is the number of classification groups of the kth classification method. The elements in C are classified according to the minimum granularity to generate a set Among them, 1, 2, ..., n4 are the numbers of the elements in the set. is the grouping interval, Among them, 1, 2, ..., n2 are the numbers of the classification intervals. For categorical data, have

[0025] Furthermore, the machine learning algorithm unit in the quality control module includes an input layer, a CNN layer, an LSTM layer, and a fully connected layer, wherein: the input layer is used to receive the grouped data transmitted from the historical data grouping unit, the CNN layer: the CNN layer is used to extract the features in the input data, and extract the local features of the parameters through the convolution operation. The convolution kernel is set to W, and the convolution operation result is H = f(W*X), where f is the CNN layer activation function, and * is the convolution operation; the LSTM network layer is used to capture the timing information in the input data. The LSTM network consists of a state unit, a forget gate, an input gate, and an output gate: the function of the state unit is to transmit the information stored in the forget gate, the input gate, and the output gate unit, and the state unit is denoted by C t , state unit C t Through the state unit C of the previous time step t-1 With the input x at that moment t To update, the state unit is updated as follows: where f t is the output of the forget gate, indicating the proportion of forgetting, i t represents the input gate, represents the proportion of new information retained, Represents the new candidate memory content, the forget gate f t The output of determines the state unit C of the previous time step t-1 The prescribed part needs to be forgotten, the mathematical formula is as follows: f t =σ(W f ·[h t-1 ,x t ]+b f ), where σ is the Sigmoid activation function, and the output value range is between 0 and 1, W f is the weight matrix of the forget gate, b f is the bias vector of the forget gate, h t-1 is the hidden state of the previous time step, x tFor the input of the current time step, add layer normalization to each layer of LSTM to speed up convergence and improve generalization ability. For the input x of the LSTM network t and the hidden state h t-1 , the calculation after layer normalization is:

[0026]

[0027] Wherein, μ and σ represent the mean and standard deviation respectively, and ∈ is a constant to prevent division by zero. The present invention dynamically adjusts the output weight of the LSTM gating unit through the adaptive function α(t) to improve the generalization ability of the model.

[0028]

[0029] Among them, T is the pharmaceutical cycle, input gate i t Determines the input x of the current time step t Need to be added to state unit C t The mathematical formula for the part is as follows: t =σ(W i ·[h t-1 ,x t ]+b c ), the new candidate memory content C t =tanh(W C ·[h t-1 ,x t ]+b C ) represents the new information added to the state unit. The mathematical formula is as follows: Where tanh is the hyperbolic tangent activation function, and the output value range is between -1 and 1. i is the weight matrix of the input gate, b i is the bias vector of the input gate, W C is the weight matrix of candidate memory content, b c is the bias vector of the candidate memory content; output gate o t The output of determines the hidden state h of the current time step t The state unit C contains the number of parts t Information, the mathematical formula is as follows: t =σ(W o ·[h t-1 ,x t ]+b o ) represents the hidden state h of the current time step t It represents the state unit information after being filtered by the output gate. The mathematical formula is as follows: t =o t tanh(C t ), where W o is the weight matrix of the output gate, bo is the bias vector of the output gate; the input of LSTM is set to the output H of the convolution layer, and through LSTM, the output H1=LSTM(H) is obtained; the fully connected layer is used to map the high-dimensional features of the convolution layer output to the prediction results, and the weight of the fully connected layer is set to V, Y=g(V·H1), where g is the activation function of the fully connected layer. For the CNN-LSTM network, the loss function is L, the weight matrix is ​​W, and the loss function of L2 regularization is expressed as:

[0030]

[0031] Among them, λ is the regularization coefficient, θ is the learning rate of the CNN-LSTM model. In order to dynamically adjust the learning rate of the CNN-LSTM model, the CNN-LSTM model introduces an adaptive learning rate adjustment algorithm based on model performance and historical gradient information. Suppose the gradient

[0032]

[0033] By time step t, calculate the loss function L(θ t ) and gradient And use exponential smoothing to update the moving average loss function value L avg,t :

[0034] L avg,t =γL avg,t-1 +(1-γ)L(θ t )

[0035] Calculate the historical exponentially weighted average of the gradient:

[0036] h t =δh t-1 +(1-δ)g t 2

[0037] Update the learning rate:

[0038] θ t+1 =θ t -α t g t

[0039] Among them, α is the parameter that controls the learning rate of the CNN-LSTM model, γ: the smoothing coefficient that controls the moving average loss function value, which is set to 0.9 in this system, δ: the smoothing coefficient that controls the influence of gradient history, which is set to 0.99 in this system, and η: the adjustment coefficient, which is used to control the sensitivity of the learning rate adjustment.

[0040] Furthermore, the system management module includes a system maintenance and monitoring unit, a security and permission control unit, a data backup and recovery unit, a log recording and auditing unit, a system update and upgrade unit, and a user support and help unit, wherein: the system maintenance and monitoring unit is used to monitor the system performance in real time, perform fault detection and diagnosis, and ensure the stable operation of the system. The system maintenance and monitoring unit can automatically identify and repair faults, and provide diagnostic reports to facilitate maintenance personnel to conduct in-depth analysis and processing; the security and permission control unit includes user identity authentication and permission control. The security and permission control unit can ensure the security of system data and prevent unauthorized access and operation. Through permission settings, the system administrator defines and assigns operation permissions to different users to ensure operation Compliance and security; data backup and recovery unit, regularly backs up system data to ensure that in the event of data loss or system crash, data can be quickly restored to reduce the impact on production; logging and auditing unit, records all system operations and event logs for easy auditing and tracking. The logging and auditing unit can record the operation behavior of each user to ensure that problems can be quickly located and resolved; system update and upgrade unit, regularly checks and updates system software to ensure that the system is always running on the latest version to obtain the latest functions and security patches; user support and help unit, provides system usage guides, technical problem answers and technical support to help users quickly get started and solve problems encountered during use, thereby improving user experience and satisfaction.

[0041] Furthermore, the system maintenance and monitoring unit constructs a Pareto function set, which includes five indicators: CPU usage, memory usage, network traffic, error log, and service score. Available computing power serves this system, and the Pareto function set Pa is expressed as:

[0042]

[0043] Among them, CPU1 is the CPU usage of the first computing power of the system, MEMORY1 is the memory usage of the first computing power of the system, STREAM1 is the network traffic of the first computing power of the system, ELOG1 is the error log of the first computing power of the system, QOS1 is the service score of the first computing power of the system, CPU2 is the CPU usage of the second computing power of the system, MEMORY2 is the memory usage of the second computing power of the system, STREAM2 is the network traffic of the second computing power of the system, ELOG2 is the error log of the second computing power of the system, QOS2 is the service score of the second computing power of the system, For the system The CPU usage of each computing power, For the system The memory usage of each computing power, For the system The network traffic of computing power, System The error log of the computing power, For the system The service score of computing power is have:

[0044] Pa v =[CPU v MEMORY v STREAM v ELOG v QOS V ]

[0045] Among them, Pa v is the Pareto function set of the vth computing power of the system, CPU v is the CPU usage of the vth computing power in the system, MEMORY v is the memory usage of the vth computing power in the system, STREAM v is the network traffic of the vth computing power in the system, ELOG v Error log of the vth computing power of the system, QOS v Score the service of the vth computing power in the system. For feature transformation, we have:

[0046]

[0047] in, After feature transformation, construct the optimal set of Pareto function Pa * ,have

[0048]

[0049] Among them, index a and index b Both The elements in, temp and temp` are index numbers, f temp and f temp` They are the temp-th objective function and the temp`-th objective function respectively, where the objective function is:

[0050]

[0051] Among them, FUNCTION is the objective function, including index and λ, where index is Pa *The elements in , λ is the constraint factor, there are λ∈{λ1,λ2,λ3,λ4,λ5}, λ1,λ2,λ3,λ4,λ5 are the constraint factors of CPU, MEMORY, STREAM, ELOG and QOS respectively, w CPU 、w MEMORY 、w STREAM 、w ELOG 、w QOS are the weight factors of CPU, MEMORY, STREAM, ELOG and QOS, respectively, CPU (index), f MEMORY (index), f STREAM (index), f ELOG (index), f QOS (index) are the sub-objective functions of CPU, MEMORY, STREAM, ELOG and QOS respectively.

[0052] The objective function has the following constraint functions:

[0053]

[0054] in, CRU v ,MEMORY v , STREAM v , ELOG v , QOS v The weight factor, λ l is the lth constraint factor, index l It is the lth group element in CPU, MEMORY, STREAM, ELOG, and QOS.

[0055] Furthermore, the data visualization and reporting module includes a data visualization unit, a historical data analysis unit, and a report generation unit, wherein: the data visualization unit provides an intuitive graphical interface to display the data in the control process. The data visualization unit can generate line graphs, bar graphs, and pie charts to help users quickly understand the changing trends and distribution of data. The real-time monitoring panel is used to display the temperature, pressure, flow rate, pH value, and dissolved oxygen content in the production process in real time. Users can monitor the production status in real time through the real-time monitoring panel and promptly discover and handle abnormal situations; the historical data analysis unit is used to analyze the historical data in the production process and generate trend graphs and statistical reports. Users can understand the long-term changing trends in the production process through the historical data analysis unit to assist in optimizing the production process; the report generation unit automatically generates an analysis report, including the changing trends of parameters, the records of abnormal events, and the processing results; the user-defined view allows the user to customize the data display interface and report content as needed. The user selects the required parameters and time period to generate personalized data views and reports to meet different analysis needs; the data export and sharing unit supports exporting visualized data and reports to word, excel, and pdf file formats.

[0056] Beneficial effects of the present invention: intelligent adjustment and optimization of the production process can be achieved, ensuring that drug purity and production efficiency reach the best state. Through the collection and analysis of historical data and real-time data, the cluster classification method can obtain the data relationship hidden in the data, the decision tree classification method can obtain the data relationship explicit in the data, and the empirical formula can obtain the conventional classification results of the data. The classification effect is optimized by dividing the classification results according to the minimum granularity. CNN is used to extract key parameter features, and time series prediction is performed through LSTM. It can accurately capture parameter changes in the production process, provide an accurate data foundation and optimization plan, and improve production efficiency and product quality. The present invention can monitor the key parameters in the production process in real time, automatically identify and adjust production conditions through machine learning algorithms, ensure the stability and consistency of the production process, avoid errors caused by human operation, and improve the degree of automation of the production process. The system's built-in model training and updating unit can regularly train and update the model, continuously improve and optimize the model with new data, adapt to changes in the production process, and ensure that the system always maintains efficient operation under different production conditions. Through monitoring and analysis of real-time data, when it is found that the key parameters in the production process are abnormal or deviate from the preset range, the system can promptly issue an early warning and feedback to the operator to ensure that measures can be taken quickly to avoid production accidents or quality problems. The system allows users to define and adjust key parameter settings according to production needs and experience, and through the optimization of machine learning algorithms, combined with historical data and real-time monitoring results, provide personalized parameter optimization solutions to improve production flexibility and efficiency. Through intelligent adjustment and optimization, the system can achieve precise control of key parameters under different production conditions, ensure the stability and consistency of drug purity and production efficiency, and meet increasingly stringent quality requirements. The present invention combines the powerful feature extraction capabilities of convolutional neural networks (CNN) and the time series prediction capabilities of long short-term memory networks (LSTM). Through an intelligent control system, it realizes the comprehensive optimization and efficient management of biopharmaceutical purification processes, providing strong technical support for the biopharmaceutical industry. By analyzing five types of indicators, namely CPU usage, memory usage, network traffic, error logs, and service scoring, a set of Pareto functions was established to achieve comprehensive monitoring and optimization of system performance. Specifically, the Pareto optimization method can identify the optimal combination of different computing services, that is, to achieve the optimal results on multiple performance indicators without losing the optimization results of other indicators, avoiding the limitations of a single indicator, and enabling the system to achieve a balance between different performance requirements, especially when system resources are limited, to ensure the optimal allocation of resources and improve the overall system operation efficiency and reliability. Through the Pareto function set, the system can dynamically adjust the resource allocation of different computing services, make full use of system resources without affecting key performance indicators, avoid resource waste, and the system automatically monitors and optimizes performance, reducing the need for manual intervention and reducing operation and maintenance costs.The intelligent purification process control system of the present invention can significantly improve the automation and intelligence level of biopharmaceutical production, improve the quality and efficiency of drug production, and meet the needs of the modern biopharmaceutical industry for efficient, safe and high-quality production. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] The invention is further illustrated by the accompanying drawings, but the embodiments in the accompanying drawings do not constitute any limitation to the invention. A person skilled in the art can obtain other drawings based on the following drawings without creative work.

[0058] Figure 1 It is a schematic diagram of the structure of the present invention. DETAILED DESCRIPTION

[0059] The present invention is further described in conjunction with the following examples.

[0060] See also Figure 1 , a biopharmaceutical intelligent purification process control system based on big data in this embodiment includes a data acquisition module, a quality control module, a system management module, and a data visualization and reporting module, wherein:

[0061] The data acquisition module is used to use sensor technology to collect parameters in the biopharmaceutical purification process and record them in the control system; the quality control module has two functions. First, it groups historical data. Second, through the historical grouped data, it uses machine learning algorithms to establish a functional relationship between the parameters in the control process and the drug purity, thereby realizing intelligent adjustment and optimization of the production process and improving drug purity; the system management module includes system maintenance and monitoring functions and security and authority control functions. The system maintenance and monitoring function plays a role in monitoring system performance and fault detection, and the security and authority control function plays a role in user identity authentication and authority control; the data visualization and reporting module provides an intuitive data visualization interface, displays the data and analysis results in the control process, and generates analysis reports.

[0062] Specifically, the data acquisition module includes a sensor technology application unit, a real-time data acquisition unit, a data recording and storage unit, a data preprocessing unit, and an interface and communication unit, wherein: the sensor technology application unit is used to monitor and collect temperature, pressure, flow rate, pH value, and dissolved oxygen content in the biopharmaceutical purification process, and the sensor is arranged inside the purification equipment; the real-time data acquisition unit is connected to the data acquisition module through a high-speed data interface to achieve a millisecond-level data acquisition frequency, and transmits the data to the central control system in real time through the data acquisition system; the data recording and storage unit has data storage capabilities, classifies and stores the collected data, adopts redundant storage technology, and regularly backs up the data; the data preprocessing unit preprocesses the data during the data acquisition process, including noise removal, outlier detection and data correction, and marking timestamps; the interface and communication unit supports docking with the quality control module, system management module, data visualization and reporting module, and connects to the remote server through a network interface to remotely monitor and analyze the data.

[0063] Specifically, the quality control module includes a historical data grouping unit, a machine learning algorithm unit, a model training and updating unit, a data feedback and early warning unit, and a user-defined parameter setting unit, wherein: the historical data grouping unit is used to group historical data according to temperature, pressure, flow rate, pH value, and dissolved oxygen content characteristics (the dissolved oxygen content is the content of oxygen dissolved in water, in mg / L) to facilitate analysis and comparison of production results and provide a data basis; the machine learning algorithm unit analyzes the historical grouping data and uses the machine learning algorithm to establish a functional relationship between the characteristics in the control process and the drug purity, so as to capture the impact of changes in data characteristics on drug purity; the intelligent adjustment and optimization unit performs real-time monitoring and intelligent adjustment of the production process based on the model established by the machine learning algorithm , the intelligent adjustment and optimization unit can automatically identify and adjust parameters; the model training and update unit can use new data to continuously improve and optimize the model, adapt to changes in the production process, and control the temperature, pressure, flow rate, pH value, and dissolved oxygen content characteristics within the preset range to accurately predict the yield and purity of biological drugs; the data feedback and early warning unit, through the monitoring and analysis of real-time data, when it is found that the parameters in the production process are abnormal and deviate from the preset range, the system can promptly issue an early warning and feedback to the operator to ensure that measures can be taken quickly to avoid production accidents and quality problems; the user-defined parameter setting unit allows the user to adjust the parameter settings according to needs, and through the optimization of machine learning algorithms, combined with historical data and real-time monitoring results, provide personalized parameter optimization solutions to improve production flexibility and efficiency.

[0064] Specifically, the historical data grouping unit in the quality control module obtains the data of the data acquisition module in the system from the database, and the module reads the historical data set. The historical data independent variable set is recorded as X = {x1, x2, x3, ..., x p ,…,x n}, the historical data dependent variable set is recorded as Y = {y1, y2, y3, ..., y p ,…,y n}, where x p ,y p Design a feature type set to purify the type. The numbers 1, 2, 3, ..., p, ..., n in the X and Y sets are the 1st, 2nd, 3rd, ..., p, ..., nth data in the historical data, denoted by x p ={xw1,xw2,xw3,xw4,x w5}, where xw1 is the temperature characteristic, xw2 is the pressure characteristic, xw3 is the flow rate characteristic, xw4 is the pH value, and xw5 is the dissolved oxygen content characteristic. p The features are regarded as a set of dependent variables, denoted as y p ={yw1,yw2}, where yw1 is the purity of the drug and yw2 is the drug yield, recorded as the classification interval set Among them, 1, 2, ..., n1 are the classification interval numbers, m is the grouping number, and k is the classification method. For categorical data, In this system, three types of classification are used: empirical formula grouping, cluster grouping, and classification tree grouping. The empirical formula grouping method is specifically as follows: determine the specific range of temperature, pressure, flow rate, pH value, and dissolved oxygen content, and set the range of temperature (T) to T min To T max , where T min is the lowest temperature of the purification process, T max The maximum temperature of the purification process is set to P. min To P max , where P min is the minimum pressure of the purification process, P max For the highest pressure of the purification process, set the flow rate (F) in the range of F min To F max , where F min is the minimum flow rate of the purification process, F max For the highest flow rate during purification, set the pH range to pH min To pH max , where pH min The lowest pH value during the purification process, pH maxFor the highest pH value of the purification process, set the dissolved oxygen content (DO) range to DO min To DO max , where DO min The minimum dissolved oxygen content in the purification process, DO max The highest dissolved oxygen content in the purification process is obtained by establishing an empirical formula S for purity prediction through the heuristic method of polynomial fitting. The data are grouped from small to large according to S, and a total of 20 groups are divided, and the group set is recorded as G m1 , where m is the group number, the maximum value is 20, and the prediction formula is as follows:

[0065] S=aT·bP·cF+d(pH)+e(DO)+r

[0066] Among them, a, b, c, d, e, and r are the parameters fitted by the heuristic algorithm. The clustering method is specifically to transform each element x in the X set i , calculate x i With the group center μ m The Euclidean distance of the center of x i Assigned to the closest group, denoted by G m2 is the set of the mth group, μ m is the center point of the m group set, μ ml is the center point of any grouping set:

[0067]

[0068] Among them, for G m2 , calculate the updated group center μ m ':

[0069]

[0070] in, G m2 The number of elements in the set is recorded. The classification tree method is specifically a gradient boosting decision tree algorithm. For each feature (temperature, pressure, flow rate, pH value, dissolved oxygen content) data, there is a classification tree model:

[0071]

[0072] Where F0(x) is the initial prediction value, γ is the leaf node prediction value, Le is the loss function, and F X is the predicted value in the X set, r nj is the model prediction residual, is the predicted value of F in the previous step,

[0073]

[0074] And fit a new decision tree, For the input x j Decision tree model:

[0075]

[0076] h X is the new decision tree obtained by fitting the model, updating the model,

[0077] F n (x) = F n-1 (x)+0.1h X (x)

[0078] F n (x) is the latest model prediction value, F n-1 (x) is F n (x) The predicted value of the previous model, the final classification result will be the set of classification methods C = {G mk ,m∈{1,2,3,…o},k∈{1,2,3}}, where 1, 2, and 3 in k represent the empirical formula method, clustering method, and classification tree method respectively, and o is the number of classification groups of the kth classification method. The elements in C are classified according to the minimum granularity to generate a set Among them, 1, 2, ..., n4 are the numbers of the elements in the set. is the grouping interval, Among them, 1, 2, ..., n2 are the numbers of the classification intervals. For categorical data, have

[0079] Specifically, the machine learning algorithm unit in the quality control module includes an input layer, a CNN layer, an LSTM layer, and a fully connected layer, wherein: the input layer is used to receive the grouped data transmitted from the historical data grouping unit, the CNN layer: the CNN layer is used to extract the features in the input data, and extract the local features of the parameters through the convolution operation. The convolution kernel is set to W, and the convolution operation result is H = f(W*X), where f is the CNN layer activation function, and * is the convolution operation; the LSTM network layer is used to capture the timing information in the input data. The LSTM network consists of a state unit, a forget gate, an input gate, and an output gate: the function of the state unit is to transmit the information stored in the forget gate, the input gate, and the output gate unit, and the state unit is denoted by C t , state unit C t Through the state unit C of the previous time step t-1 With the input x at that moment t To update, the state unit is updated as follows: where f tis the output of the forget gate, indicating the proportion of forgetting, i t represents the input gate, represents the proportion of new information retained, Represents the new candidate memory content, the forget gate f t The output of determines the state unit C of the previous time step t-1 The prescribed part needs to be forgotten, the mathematical formula is as follows: f t =σ(W f ·[h t-1 ,x t ]+b f ), where σ is the Sigmoid activation function, and the output value range is between 0 and 1, W f is the weight matrix of the forget gate, b f is the bias vector of the forget gate, h t-1 is the hidden state of the previous time step, x t For the input of the current time step, add layer normalization to each layer of LSTM to speed up convergence and improve generalization ability. For the input x of the LSTM network t and the hidden state h t-1 , the calculation after layer normalization is:

[0080]

[0081] Wherein, μ and σ represent the mean and standard deviation respectively, and ∈ is a constant to prevent division by zero. The present invention dynamically adjusts the output weight of the LSTM gating unit through the adaptive function α(t) to improve the generalization ability of the model.

[0082]

[0083] Among them, T is the pharmaceutical cycle, input gate i t Determines the input x of the current time step t Need to be added to state unit C t The mathematical formula for the part is as follows: t =σ(W i ·[h t-1 ,x t ]+b c ), the new candidate memory content C t =tanh(W C ·[h t-1 ,x t ]+b C ) represents the new information added to the state unit. The mathematical formula is as follows: Where tanh is the hyperbolic tangent activation function, and the output value range is between -1 and 1. i is the weight matrix of the input gate, b i is the bias vector of the input gate, W Cis the weight matrix of candidate memory content, b c is the bias vector of the candidate memory content; output gate o t The output of determines the hidden state h of the current time step t The state unit C contains the number of parts t Information, the mathematical formula is as follows: t =σ(W o ·[h t-1 ,x t ]+b o ) represents the hidden state h of the current time step t It represents the state unit information after being filtered by the output gate. The mathematical formula is as follows: t =o t tanh(C t ), where W o is the weight matrix of the output gate, b o is the bias vector of the output gate; the input of LSTM is set to the output H of the convolution layer, and through LSTM, the output H1=LSTM(H) is obtained; the fully connected layer is used to map the high-dimensional features of the convolution layer output to the prediction results, and the weight of the fully connected layer is set to V, Y=g(V·H1), where g is the activation function of the fully connected layer. For the CNN-LSTM network, the loss function is L, the weight matrix is ​​W, and the loss function of L2 regularization is expressed as:

[0084]

[0085] Among them, λ is the regularization coefficient, θ is the learning rate of the CNN-LSTM model. In order to dynamically adjust the learning rate of the CNN-LSTM model, the CNN-LSTM model introduces an adaptive learning rate adjustment algorithm based on model performance and historical gradient information. Suppose the gradient

[0086]

[0087] By time step t, calculate the loss function L(θ t ) and gradient And use exponential smoothing to update the moving average loss function value L avg,t :

[0088] L avg,t =γL avg,t-1 +(1-γ)L(θ t )

[0089] Calculate the historical exponentially weighted average of the gradient:

[0090] h t =δh t-1 +(1-δ)g t 2

[0091] Update the learning rate:

[0092] θ t+1 =θ t -α t g t

[0093] Among them, α is the parameter that controls the learning rate of the CNN-LSTM model, γ: the smoothing coefficient that controls the moving average loss function value, which is set to 0.9 in this system, δ: the smoothing coefficient that controls the influence of gradient history, which is set to 0.99 in this system, and η: the adjustment coefficient, which is used to control the sensitivity of the learning rate adjustment.

[0094] Specifically, the system management module includes a system maintenance and monitoring unit, a security and permission control unit, a data backup and recovery unit, a log recording and auditing unit, a system update and upgrade unit, and a user support and help unit, wherein: the system maintenance and monitoring unit is used to monitor the system performance in real time, perform fault detection and diagnosis, and ensure the stable operation of the system. The system maintenance and monitoring unit can automatically identify and repair faults, and provide diagnostic reports to facilitate maintenance personnel to conduct in-depth analysis and processing; the security and permission control unit includes user identity authentication and permission control. The security and permission control unit can ensure the security of system data and prevent unauthorized access and operation. Through permission settings, the system administrator defines and assigns operation permissions to different users to ensure operation Compliance and security; data backup and recovery unit, regularly backs up system data to ensure that in the event of data loss or system crash, data can be quickly restored to reduce the impact on production; logging and auditing unit, records all system operations and event logs for easy auditing and tracking. The logging and auditing unit can record the operation behavior of each user to ensure that problems can be quickly located and resolved; system update and upgrade unit, regularly checks and updates system software to ensure that the system is always running on the latest version to obtain the latest functions and security patches; user support and help unit, provides system usage guides, technical problem answers and technical support to help users quickly get started and solve problems encountered during use, thereby improving user experience and satisfaction.

[0095] Specifically, the system maintenance and monitoring unit constructs a Pareto function set, which includes five indicators: CPU usage, memory usage, network traffic, error log, and service score. Available computing power serves this system, and the Pareto function set Pa is expressed as:

[0096]

[0097] Among them, CPU1 is the CPU usage of the first computing power of the system, MEMORY1 is the memory usage of the first computing power of the system, STRAM1 is the network traffic of the first computing power of the system, ELOG1 is the error log of the first computing power of the system, QOS1 is the service score of the first computing power of the system, CPU2 is the CPU usage of the second computing power of the system, MEMORY2 is the memory usage of the second computing power of the system, STRAM2 is the network traffic of the second computing power of the system, ELOG2 is the error log of the second computing power of the system, QOS2 is the service score of the second computing power of the system, For the system The CPU usage of each computing power, For the system The memory usage of each computing power, For the system The network traffic of computing power, System The error log of the computing power, For the system The service score of computing power is have:

[0098] Pa v =[CPU v MEMORY v STREAM v ELOG v QOS V ]

[0099] Among them, Pa v is the Pareto function set of the vth computing power of the system, CPU v is the CPU usage of the vth computing power in the system, MEMORY v is the memory usage of the vth computing power in the system, STREAM v is the network traffic of the vth computing power in the system, ELOG v Error log of the vth computing power of the system, QOS v Score the service of the vth computing power in the system. For feature transformation, we have:

[0100]

[0101] in, After feature transformation, construct the optimal set of Pareto function Pa * ,have

[0102]

[0103] Among them, index a and indexb Both The elements in, temp and temp` are index numbers, f temp and f temp` They are the temp-th objective function and the temp`-th objective function respectively, where the objective function is:

[0104]

[0105] Among them, FUNCTION is the objective function, including index and λ, where index is Pa * The elements in , λ is the constraint factor, there are λ∈{λ1,λ2,λ3,λ4,λ5}, λ1,λ2,λ3,λ4,λ5 are the constraint factors of CPU, MEMORY, STREAM, ELOG and QOS respectively, w CPU 、w MEMORY 、w STREAM 、w ELOG 、w QOS are the weight factors of CPU, MEMORY, STREAM, ELOG and QOS, respectively, CPU (index), f MEMORY (index), f STREAM (index), f ELOG (index), f QOS (index) are the sub-objective functions of CPU, MEMORY, STREAM, ELOG and QOS respectively.

[0106] The objective function has the following constraint functions:

[0107]

[0108] in, CPU v ,MEMORY v , STREAM v , ELOG v , QOS v The weight factor, λ l is the lth constraint factor, index lIt is the lth group element in CPU, MEMORY, STREAM, ELOG, and QOS. Specifically, the data visualization and reporting module includes a data visualization unit, a historical data analysis unit, and a report generation unit, wherein: the data visualization unit provides an intuitive graphical interface to display the data in the control process. The data visualization unit can generate line graphs, bar graphs, and pie charts to help users quickly understand the changing trends and distribution of data. The real-time monitoring panel is used to display the temperature, pressure, flow rate, pH value, and dissolved oxygen content in the production process in real time. Users can monitor the production status in real time through the real-time monitoring panel and promptly discover and handle abnormal situations; the historical data analysis unit is used to analyze the historical data in the production process and generate trend graphs and statistical reports. Users can understand the long-term changing trends in the production process through the historical data analysis unit to assist in optimizing the production process; the report generation unit automatically generates an analysis report, including the changing trends of parameters, the records of abnormal events, and the processing results; the user-defined view allows the user to customize the data display interface and report content as needed. The user selects the required parameters and time period to generate personalized data views and reports to meet different analysis needs; the data export and sharing unit supports exporting visualized data and reports to word, excel, and pdf file formats.

[0109] Beneficial effects of the present invention: intelligent adjustment and optimization of the production process can be achieved, ensuring that drug purity and production efficiency reach the best state. Through the collection and analysis of historical data and real-time data, the cluster classification method can obtain the data relationship hidden in the data, the decision tree classification method can obtain the data relationship explicit in the data, and the empirical formula can obtain the conventional classification results of the data. The classification effect is optimized by dividing the classification results according to the minimum granularity. CNN is used to extract key parameter features, and time series prediction is performed through LSTM. It can accurately capture parameter changes in the production process, provide an accurate data foundation and optimization plan, and improve production efficiency and product quality. The present invention can monitor the key parameters in the production process in real time, automatically identify and adjust production conditions through machine learning algorithms, ensure the stability and consistency of the production process, avoid errors caused by human operation, and improve the degree of automation of the production process. The system's built-in model training and updating unit can regularly train and update the model, continuously improve and optimize the model with new data, adapt to changes in the production process, and ensure that the system always maintains efficient operation under different production conditions. Through monitoring and analysis of real-time data, when it is found that the key parameters in the production process are abnormal or deviate from the preset range, the system can promptly issue an early warning and feedback to the operator to ensure that measures can be taken quickly to avoid production accidents or quality problems. The system allows users to define and adjust key parameter settings according to production needs and experience, and through the optimization of machine learning algorithms, combined with historical data and real-time monitoring results, provide personalized parameter optimization solutions to improve production flexibility and efficiency. Through intelligent adjustment and optimization, the system can achieve precise control of key parameters under different production conditions, ensure the stability and consistency of drug purity and production efficiency, and meet increasingly stringent quality requirements. The present invention combines the powerful feature extraction capabilities of convolutional neural networks (CNN) and the time series prediction capabilities of long short-term memory networks (LSTM). Through an intelligent control system, it realizes the comprehensive optimization and efficient management of biopharmaceutical purification processes, providing strong technical support for the biopharmaceutical industry. In the paper, by analyzing five types of indicators, namely CPU usage, memory usage, network traffic, error logs, and service scores, a set of Pareto functions was established to achieve comprehensive monitoring and optimization of system performance. Specifically, the Pareto optimization method can identify the optimal combination of different computing power services, that is, to achieve the optimal optimization results on multiple performance indicators at the same time without losing other indicators, avoiding the limitations of a single indicator, and enabling the system to achieve a balance between different performance requirements, especially when system resources are limited, to ensure the optimal allocation of resources and improve the overall system operation efficiency and reliability. Through the Pareto function set, the system can dynamically adjust the resource allocation of different computing power services, make full use of system resources without affecting key performance indicators, avoid resource waste, and the system automatically monitors and optimizes performance, reducing the need for manual intervention and reducing operation and maintenance costs.The intelligent purification process control system of the present invention can significantly improve the automation and intelligence level of biopharmaceutical production, improve the quality and efficiency of drug production, and meet the needs of the modern biopharmaceutical industry for efficient, safe and high-quality production.

[0110] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention, rather than to limit the scope of protection of the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the essence and scope of the technical solution of the present invention.

Claims

1. A biopharmaceutical purification process control system based on big data, including the following modules: a data acquisition module, which is used to collect parameters in the biopharmaceutical purification process using sensor technology and record them in the control system; The quality control module has two functions. First, it groups historical data. Second, it uses machine learning algorithms to establish a functional relationship between parameters in the control process and drug purity through historical grouped data, so as to achieve intelligent adjustment and optimization of the production process and improve drug purity. The system management module includes system maintenance and monitoring functions and security and authority control functions. The system maintenance and monitoring function plays a role in monitoring system performance and fault detection, and the security and authority control function plays a role in user identity authentication and authority control. Data visualization and reporting module, which provides an intuitive data visualization interface, displays the data and analysis results of the control process, and generates analysis reports; The quality control module includes a historical data grouping unit, a machine learning algorithm unit, a model training and updating unit, a data feedback and early warning unit, and a user-defined parameter setting unit, wherein: the historical data grouping unit is used to group historical data according to temperature, pressure, flow rate, pH value, and dissolved oxygen content characteristics, wherein the dissolved oxygen content is the content of oxygen dissolved in water, in mg / L, so as to facilitate analysis and comparison of production results and provide a data basis; the machine learning algorithm unit analyzes historical grouping data and uses machine learning algorithms to establish a functional relationship between the characteristics in the control process and the drug purity, and can capture the impact of changes in data characteristics on drug purity; the intelligent adjustment and optimization unit monitors the production process in real time and makes intelligent adjustments based on the model established by the machine learning algorithm, and intelligently adjusts the production process based on the model established by the machine learning algorithm. The adjustment and optimization unit can automatically identify and adjust parameters; the model training and update unit can use new data to continuously improve and optimize the model, adapt to changes in the production process, and control the temperature, pressure, flow rate, pH value, and dissolved oxygen content characteristics within the preset range to accurately predict the yield and purity of biological drugs; the data feedback and early warning unit, through the monitoring and analysis of real-time data, when it is found that the parameters in the production process are abnormal and deviate from the preset range, the system can promptly issue an early warning and feedback to the operator to ensure that measures can be taken quickly to avoid production accidents and quality problems; the user-defined parameter setting unit allows users to adjust parameter settings according to needs, and through the optimization of machine learning algorithms, combined with historical data and real-time monitoring results, provides personalized parameter optimization solutions to improve production flexibility and efficiency; The historical data grouping unit is specifically as follows: the data of the data acquisition module in the system is obtained from the database, the module reads the historical data set, and the historical data independent variable set is recorded as X = {x1, x2, x3, ..., x p ,…,x n }, the historical data dependent variable set is recorded as Y = {y1, y2, y3, ..., y p ,…,y n }, where x p ,y p Design a feature type set to purify the type. The numbers 1, 2, 3, ..., p, ..., n in the X and Y sets are the 1st, 2nd, 3rd, ..., p, ..., nth data in the historical data, denoted by x p ={xw1,xw2,xw3,xw4,x w5 }, where xw1 is the temperature characteristic, xw2 is the pressure characteristic, xw3 is the flow rate characteristic, xw4 is the pH value, and xw5 is the dissolved oxygen content characteristic. p The features are regarded as a set of dependent variables, denoted as y p ={yw1,yw2}, where yw1 is the purity of the drug and yw2 is the drug yield, recorded as the classification interval set Among them, 1, 2, ..., n1 are the classification interval numbers, m is the grouping number, and k is the classification method. For categorical data, In this system, three types of classification are used: empirical formula grouping, cluster grouping, and classification tree grouping. The empirical formula grouping method is specifically as follows: determine the specific range of temperature, pressure, flow rate, pH value, and dissolved oxygen content, and set the range of temperature (T) to T min To T max , where T min is the lowest temperature of the purification process, T max The maximum temperature of the purification process is set to P. min To P max , where P min is the minimum pressure of the purification process, P max For the highest pressure of the purification process, set the flow rate (F) in the range of F min To F max , where F min is the minimum flow rate of the purification process, F max For the highest flow rate during purification, set the pH range to pH min To pH max , where pH min The lowest pH value during the purification process, pH max For the highest pH value of the purification process, set the dissolved oxygen content (DO) range to DO min To DO max , where DO min The minimum dissolved oxygen content in the purification process, DO max The highest dissolved oxygen content in the purification process is obtained by establishing an empirical formula S for purity prediction through the heuristic method of polynomial fitting. The data are grouped from small to large according to S, and a total of 20 groups are divided, and the group set is recorded as G m1 , where m is the group number, the maximum value is 20, and the prediction formula is as follows: S=aT·bP·cF+d(pH)+e(DO)+r Among them, a, b, c, d, e, and r are the parameters fitted by the heuristic algorithm. The clustering method is to transform each element x in the X set into i , calculate x i With the group center μ m The Euclidean distance of the center of x i Assigned to the closest group, denoted by G m2 is the set of the mth group, μ m is the center point of the m group set, μ ml is the center point of any grouping set: Among them, for G m2 , calculate the updated group center μ m ': in, G m2 The number of elements in the set is recorded. The classification tree method is specifically a gradient boosting decision tree algorithm. For each feature (temperature, pressure, flow rate, pH value, dissolved oxygen content) data, there is a classification tree model: F0(x)=argmin γ S j Le(y j ,c) Where F0(x) is the initial prediction value, γ is the leaf node prediction value, Le is the loss function, and F X is the predicted value in the X set, r nj is the model prediction residual, is the predicted value of F in the previous step, And fit a new decision tree, For the input x j Decision tree model: h X is the new decision tree obtained by fitting the model, updating the model, F n (x)=F n-1 (x)+0.1h X (x) F n (x) is the latest model prediction value, F n-1 (x) is F n (x) The predicted value of the previous model, the final classification result will be the set of classification methods C = {G mk ,m∈{1,2,3,…o},k∈{1,2,3}}, where 1, 2, and 3 in k represent the empirical formula method, clustering method, and classification tree method respectively, and o is the number of classification groups of the kth classification method. The elements in C are classified according to the minimum granularity to generate a set Among them, 1, 2, ..., n4 are the numbers of the elements in the set. is the grouping interval, Among them, 1, 2, ..., n2 are the numbers of the classification intervals. For categorical data, have 2. A biopharmaceutical purification process control system based on big data according to claim 1, characterized in that: The data acquisition module includes a sensor technology application unit, a real-time data acquisition unit, a data recording and storage unit, a data preprocessing unit, and an interface and communication unit, wherein: the sensor technology application unit is used to monitor and collect temperature, pressure, flow rate, pH value, and dissolved oxygen content in the biopharmaceutical purification process, and the sensor is arranged inside the purification equipment; the real-time data acquisition unit is connected to the data acquisition module through a high-speed data interface to achieve a millisecond-level data acquisition frequency, and transmits the data to the central control system in real time through the data acquisition system; the data recording and storage unit has data storage capabilities, classifies and stores the collected data, adopts redundant storage technology, and regularly backs up the data; the data preprocessing unit preprocesses the data during the data acquisition process, including noise removal, outlier detection and data correction, and marking timestamps; the interface and communication unit supports docking with the quality control module, the system management module, and the data visualization and reporting module, and connects to the remote server through the network interface to remotely monitor and analyze the data.

3. A biopharmaceutical purification process control system based on big data according to claim 1, characterized in that: The machine learning algorithm unit includes an input layer, a CNN layer, an LSTM layer, and a fully connected layer, wherein: the input layer is used to receive the grouped data transmitted from the historical data grouping unit; the CNN layer: the CNN layer is used to extract features in the input data, extract local features of the parameters through convolution operations, set the convolution kernel to W, and the convolution operation result is H=f(W*X), wherein f is the CNN layer activation function, and * is the convolution operation; the LSTM network layer is used to capture the timing information in the input data, and the LSTM network is composed of a state unit, a forget gate, an input gate, and an output gate: wherein the function of the state unit is to transmit the information stored in the forget gate, the input gate, and the output gate units, and the state unit is denoted as C t , state unit C t Through the state unit C of the previous time step t-1 With the input x at that moment t To update, the state unit is updated as follows: where f t is the output of the forget gate, indicating the proportion of forgetting, i t represents the input gate, represents the proportion of new information retained, Represents the new candidate memory content, the forget gate f t The output of determines the state unit C of the previous time step t-1 The prescribed part needs to be forgotten, the mathematical formula is as follows: f t =σ(W f ·[h t-1 ,x t ]+b f ), where σ is the Sigmoid activation function, and the output value range is between 0 and 1, W f is the weight matrix of the forget gate, b f is the bias vector of the forget gate, h t-1 is the hidden state of the previous time step, x t For the input of the current time step, add layer normalization to each layer of LSTM to speed up convergence and improve generalization ability. For the input x of the LSTM network t and the hidden state h t-1 , the calculation after layer normalization is: Wherein, μ and σ represent the mean and standard deviation respectively, and ∈ is a constant to prevent division by zero. The present invention dynamically adjusts the output weight of the LSTM gating unit through the adaptive function α(t) to improve the generalization ability of the model. Among them, T is the pharmaceutical cycle, input gate i t Determines the input x of the current time step t Need to be added to state unit C t The mathematical formula for the part is as follows: t =σ(W i ·[h t-1 ,x t ]+b c ), the new candidate memory content C t =tanh(W C ·[h t-1 ,x t ]+b C ) represents the new information added to the state unit. The mathematical formula is as follows: Where tanh is the hyperbolic tangent activation function, and the output value range is between -1 and 1. i is the weight matrix of the input gate, b i is the bias vector of the input gate, W C is the weight matrix of candidate memory content, b c is the bias vector of the candidate memory content; output gate o t The output of determines the hidden state h of the current time step t The state unit C contains the number of parts t Information, the mathematical formula is as follows: t =σ(W o ·[h t-1 ,x t ]+b o ) represents the hidden state h of the current time step t It represents the state unit information after being filtered by the output gate. The mathematical formula is as follows: t =o t tanh(C t ), where W o is the weight matrix of the output gate, b o is the bias vector of the output gate; the input of LSTM is set to the output H of the convolution layer, and through LSTM, the output H1=LSTM(H) is obtained; the fully connected layer is used to map the high-dimensional features of the convolution layer output to the prediction results, and the weight of the fully connected layer is set to V, Y=g(V·H1), where g is the activation function of the fully connected layer. For the CNN-LSTM network, the loss function is L, the weight matrix is ​​W, and the loss function of L2 regularization is expressed as: Among them, λ is the regularization coefficient, θ is the learning rate of the CNN-LSTM model. In order to dynamically adjust the learning rate of the CNN-LSTM model, the CNN-LSTM model introduces an adaptive learning rate adjustment algorithm based on model performance and historical gradient information. Suppose the gradient By time step t, the loss function L(θ t ) and gradient And use exponential smoothing to update the moving average loss function value L avg,t : L avg,t =γL avg,t-1 +(1-γ)L(θ t ) Calculate the historical exponentially weighted average of the gradient: h t =δh t-1 +(1-δ)g t 2 Update the learning rate: i t+1 =θ t -a t g t Among them, α is the parameter that controls the learning rate of the CNN-LSTM model, γ: the smoothing coefficient that controls the moving average loss function value, which is set to 0.9 in this system, δ: the smoothing coefficient that controls the influence of gradient history, which is set to 0.99 in this system, and η: the adjustment coefficient, which is used to control the sensitivity of the learning rate adjustment.

4. A biopharmaceutical purification process control system based on big data according to claim 1, characterized in that: The system management module includes a system maintenance and monitoring unit, a security and authority control unit, a data backup and recovery unit, a log recording and auditing unit, a system update and upgrade unit, and a user support and help unit, wherein: the system maintenance and monitoring unit is used to monitor system performance in real time, perform fault detection and diagnosis, and ensure the stable operation of the system. The system maintenance and monitoring unit can automatically identify and repair faults, and provide diagnostic reports to facilitate maintenance personnel to conduct in-depth analysis and processing; the security and authority control unit includes user identity authentication and authority control. The security and authority control unit can ensure the security of system data and prevent unauthorized access and operation. Through authority settings, the system administrator defines and allocates operation permissions for different users to ensure the legality of operations. Compliance and security; data backup and recovery unit, regularly backs up system data to ensure that in the event of data loss or system crash, data can be quickly restored to reduce the impact on production; logging and auditing unit, records all system operations and event logs for easy auditing and tracking. The logging and auditing unit can record the operation behavior of each user to ensure that problems can be quickly located and resolved; system update and upgrade unit, regularly checks and updates system software to ensure that the system is always running on the latest version to obtain the latest functions and security patches; user support and help unit, provides system usage guides, technical problem answers and technical support to help users quickly get started and solve problems encountered during use, thereby improving user experience and satisfaction.

5. A biopharmaceutical purification process control system based on big data according to claim 4, characterized in that: The system maintenance and monitoring unit constructs a Pareto function set, which includes five indicators: CPU usage, memory usage, network traffic, error log, and service score. Available computing power serves this system, and the Pareto function set Pa is expressed as: Among them, CPU1 is the CPU usage of the first computing power of the system, MEMORY1 is the memory usage of the first computing power of the system, STREAM1 is the network traffic of the first computing power of the system, ELOG1 is the error log of the first computing power of the system, QOS1 is the service score of the first computing power of the system, CPU2 is the CPU usage of the second computing power of the system, MEMORY2 is the memory usage of the second computing power of the system, STREAM2 is the network traffic of the second computing power of the system, ELOG2 is the error log of the second computing power of the system, QOS2 is the service score of the second computing power of the system, For the system The CPU usage of each computing power, For the system The memory usage of each computing power, For the system The network traffic of computing power, System The error log of the computing power, For the system The service score of computing power is have: Pa v =[CPU v MEMORY v STREAM v ELOG v QOS V ] Among them, Pa v is the Pareto function set of the vth computing power of the system, CPU v is the CPU usage of the vth computing power in the system, MEMORY v is the memory usage of the vth computing power in the system, STREAM v is the network traffic of the vth computing power in the system, ELOG v Error log of the vth computing power of the system, QOS v Score the service of the vth computing power in the system. For feature transformation, we have: in, After feature transformation, construct the optimal set of Pareto function Pa * ,have Among them, index a and index b Both The elements in, temp and temp` are index numbers, f temp and f temp` They are the temp-th objective function and the temp`-th objective function respectively, where the objective function is: Among them, FUNCTION is the objective function, including index and λ, where index is Pa * The elements in , λ is the constraint factor, there are λ∈{λ1,λ2,λ3,λ4,λ5}, λ1,λ2,λ3,λ4,λ5 are the constraint factors of CPU, MEMORY, STREAM, ELOG and QOS respectively, w GPU 、w MEMORY 、w STREAM 、w ELOG 、w QOS are the weight factors of CPU, MEMORY, STREAM, ELOG and QOS, respectively, CPU (index), f MEMORY (index), f STREAM (index), f ELOG (index), f QOS (index) are the sub-objective functions of CPU, MEMORY, STREAM, ELOG and QOS respectively. The objective function has the following constraint functions: in, CPU v ,MEMORY v , STREAM v , ELOG v , QOS v The weight factor, λ l is the lth constraint factor, index l It is the lth group element in CPU, MEMORY, STREAM, ELOG, and QOS.

6. A biopharmaceutical purification process control system based on big data according to claim 1, characterized in that: The data visualization and reporting module includes a data visualization unit, a historical data analysis unit, and a report generation unit, wherein: the data visualization unit provides an intuitive graphical interface to display the data in the control process. The data visualization unit can generate line graphs, bar graphs, and pie charts to help users quickly understand the changing trends and distribution of data. The real-time monitoring panel is used to display the temperature, pressure, flow rate, pH value, and dissolved oxygen content in the production process in real time. Users can monitor the production status in real time through the real-time monitoring panel and promptly discover and handle abnormal situations; the historical data analysis unit is used to analyze the historical data in the production process and generate trend graphs and statistical reports. Users can understand the long-term changing trends in the production process through the historical data analysis unit to assist in optimizing the production process; the report generation unit automatically generates an analysis report, including the changing trends of parameters, the records of abnormal events, and the processing results; the user-defined view allows the user to customize the data display interface and report content as needed. The user selects the required parameters and time period to generate personalized data views and reports to meet different analysis needs; the data export and sharing unit supports exporting visualized data and reports into word, excel, and pdf file formats.

Images