Dry granulation process parameter AI optimal regulation and control method and robot system

Abstract

The invention relates to the technical field of granulation process parameter regulation and control, in particular to a dry granulation process parameter AI optimal regulation and control method and a robot system. The regulation and control method comprises the following steps: S1, parameter identification: identifying key process parameters in a dry granulation process; s2, parameter acquisition: acquiring key process parameters and quality attribute parameters in a dry granulation process, and forming a multi-dimensional data lake; s3, constructing a mapping model; s4, parameter optimization: according to the mapping model, calculating an optimization interval of the key process parameter corresponding to the required quality attribute parameter; and S5, regulation and control decision and action: acquiring real-time key process parameters, judging whether each parameter in the key process parameters is in the corresponding optimization interval, and if not, regulating the key process parameters. According to the invention, automatic parameter adjustment and optimization of the control system based on AI are realized, and the problems of untimely parameter adjustment, low process quality control level, low efficiency, high dependency on artificial experience and the like in the traditional production process are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of granulation process parameter control, and in particular to an AI optimal control method for dry granulation process parameters and a robot system. Background Art

[0002] Dry granulation is a critical step in solid dosage form production, with granulation quality directly impacting subsequent tableting, filling, and dissolution in the finished product. Excessively large particle size can affect tablet hardness and dissolution, while too small can lead to poor flowability and static electricity accumulation. Excessively high moisture content can lead to agglomeration, while too low moisture content can cause granules to fracture. Therefore, monitoring and controlling critical quality attributes (CQAs) such as particle size distribution and moisture content during dry granulation can help improve batch-to-batch consistency in drug quality.

[0003] Currently, the production process is highly dependent on human experience, with the industry still primarily relying on a closed loop of "manual sampling - laboratory testing - empirical interpretation - manual parameter adjustment." The granulation process primarily relies on manual sampling, visual inspection, or offline endpoint testing. Operators adjust process parameters based on their experience. This approach presents significant bottlenecks: First, manual sampling and laboratory testing are time-consuming, leading to delayed process parameter adjustments and slowing production. Second, operator experience varies widely, and empirical judgments lack quantitative indicators, potentially leading to quality deviations.

[0004] In addition, due to the real-time nature of measurement and control, manual judgment cannot respond to fluctuations in material properties or variations in process parameters in real time. The control logic is rigid, and process parameters such as environmental parameters, inlet air flow, inlet air temperature, material temperature, outlet air temperature, and atomization pressure are manually adjusted or set based on experience. This involves the coupling relationship of multiple variables. The commonly used single-loop or fixed PID algorithm cannot adapt to changes in quality attributes caused by parameter interactions, resulting in quality fluctuations and large differences between batches. Summary of the Invention

[0005] To address the aforementioned technical issues in the prior art, the present invention provides an AI-based optimal control method and robotic system for dry granulation process parameters. By online monitoring of process parameters and quality attributes, real-time feedback control of process parameters is implemented, replacing traditional manual sampling and empirical adjustment of process parameters. This overcomes issues such as untimely process parameter adjustment, low process quality control, low efficiency, and high reliance on manual experience, thereby enabling precise control of pharmaceutical process quality and improving quality consistency.

[0006] To achieve the above object, the technical solution of the present invention is as follows:

[0007] In a first aspect, the present invention provides a method for optimally controlling the parameters AI of a dry granulation process, comprising:

[0008] S1. Parameter identification: Identify key process parameters during dry granulation;

[0009] S2. Parameter acquisition: Acquire key process parameters and quality attribute parameters during the dry granulation process. Establish a correspondence between key process parameters and quality attribute parameters in a time series manner, and combine multiple sets of key process parameters and quality attribute parameters into a multi-dimensional data lake.

[0010] S3. Build a mapping model: Use convolutional neural network modeling to establish a mapping model between key process parameters and quality attribute parameters;

[0011] S4. Parameter optimization: Calculating the optimization intervals of key process parameters corresponding to the required quality attribute parameters based on the mapping model;

[0012] S5. Control decisions and actions: Obtain real-time key process parameters, and based on the optimization range of the key process parameters obtained in S4, determine whether each of the key process parameters is within the corresponding optimization range. If not, adjust the key process parameters until each of the key process parameters is within the corresponding optimization range.

[0013] Furthermore, in step S2, multiple sets of key process parameters and quality attribute parameters are combined into a multi-dimensional data lake, specifically including:

[0014] Key process parameters and quality attribute parameters are synchronized, outliers are corrected, and missing values ​​are filled; then all data are aligned and fused to obtain a multi-dimensional data lake.

[0015] Furthermore, in step S3, a convolutional neural network is used to construct a mapping model between key process parameters and quality attribute parameters. The specific method includes:

[0016] S301, input data representation:

[0017] The key process parameters of the input are expressed as a vector X:

[0018] X=[X i ]

[0019] Among them, X i represents the i-th key process parameter;

[0020] S302, convolutional layer output:

[0021] After the convolution operation, the output feature map F out , feature map F out Characterize the relationship between key process parameters and quality attribute parameters:

[0022]

[0023] Among them, X in (m, n) represents the key process parameters of the input, K(m, n) is the convolution kernel, and f is the activation function; (m, n) represents the spatial coordinates of the feature map;

[0024] S303, output mapping relationship:

[0025] The feature map F is transformed into out Mapped to quality attribute parameters; the output quality attribute parameter formula is:

[0026] [Y j ]=W·F out +b

[0027] Among them, Y j represents the j-th quality attribute parameter; W represents the weight matrix, and b is the bias term.

[0028] Furthermore, the key process parameters include atomization pressure, ambient temperature and humidity, relative density of incoming materials, inlet air temperature, inlet air flow, extract peristaltic pump infusion frequency, material temperature, and outlet air temperature.

[0029] Furthermore, quality attribute parameters include particle size distribution and particle moisture.

[0030] Furthermore, the activation function f is a ReLU function.

[0031] Furthermore, in step S4, based on the mapping model, a non-dominated sorting genetic algorithm is used to calculate the optimization interval of the key process parameters corresponding to the required quality attribute parameters; specifically, the following steps are performed:

[0032] S401, initialization:

[0033] Randomly generate parent population P t , each individual is a set of key process parameter vectors;

[0034] S402, fast non-dominated sorting and congestion calculation:

[0035] P t Perform fast non-dominated sorting to obtain several non-dominated frontiers F1, F2, ...;

[0036] The individual crowding distance is calculated within each frontier to maintain diversity;

[0037] S403, Selection, Crossover and Mutation:

[0038] The sire pair was selected using the tournament selection method;

[0039] Perform crossover operator to generate offspring Q t ;

[0040] Apply the mutation operator to the offspring;

[0041] S404, Merge and New Generation:

[0042] Merge R t =P t ∪Q t ;

[0043] R t Perform non-dominated sorting again and select the first M individuals to form P according to the frontier and congestion degree. t+1 ;

[0044] S405, termination and output:

[0045] When the maximum algebra G is reached max Or when the population converges, the Pareto optimal solution set composed of all individuals in the first non-dominated frontier F1 and its corresponding parameter interval are output.

[0046] Furthermore, the AI ​​optimal control method for dry granulation process parameters also includes:

[0047] S6. Model update: establishing a corresponding relationship between the key process parameters adjusted in S5 and the adjusted real-time quality attribute parameters, and updating them into the mapping model.

[0048] Furthermore, in step S2, the quality attribute parameters in the dry granulation process are obtained by:

[0049] A near-infrared spectral camera is used to obtain a spectral image of the particle sample, and a white light camera is used to obtain a white light image of the particle sample. The particle size distribution and particle moisture of the particle sample are predicted based on the spectral image and the white light image.

[0050] Furthermore, the particle size distribution and particle moisture of the particle sample are predicted based on the spectral image and the white light image, specifically including:

[0051] Predict particle size distribution and particle moisture based on spectral images,

[0052] The specific formula for particle size distribution prediction is:

[0053] D 1P =f NIR (S NIR )

[0054] Among them, D 1P is the particle size distribution predicted by the spectral image, S NIR is the spectral feature of the near-infrared spectrum image, f NIR is the function of the regression model to predict the particle size distribution based on the spectral characteristics;

[0055] The specific formula for particle moisture prediction is:

[0056] H moisture =g NIR (S NIR )

[0057] Among them, H moisture is the particle moisture, S NIR is the spectral feature of the near-infrared spectral image, g NIR is the function of the regression model to predict the particle moisture based on the spectral characteristics;

[0058] The particle size distribution is predicted based on white light images. The specific formula is:

[0059] D 2P =h WL (I WL )

[0060] Among them, D 2P is the particle size distribution predicted by white light imaging, I WL is the geometric feature of the white light image, h WL It is the function of the regression model to predict the particle size distribution based on the geometric characteristics of the white light image;

[0061] The final particle size distribution is:

[0062] D P =α·D 1P +β·D 2P

[0063] Among them, α and β are weight coefficients, α+β=1.

[0064] Furthermore, the gradient descent method is used to determine the weight coefficients α and β, and the objective function is defined as

[0065]

[0066] Where N represents the total sample size, represents the true particle size distribution of the i-th sample, D 1P (i) represents the particle size distribution of the i-th sample predicted by the spectral image, D 2P (i) represents the particle size distribution of the i-th sample predicted by the white light image;

[0067] Use the gradient descent method to minimize J(α, β) and obtain the weight coefficients α and β.

[0068] Furthermore, the update rule of the gradient descent method is:

[0069]

[0070] Among them, α' and β' are the updated weight coefficients, and η is the learning rate.

[0071] In a second aspect, the present invention also provides an AI optimal control robot system for dry granulation process parameters, comprising:

[0072] A multimodal intelligent perception module is used to obtain key process parameters and quality attribute parameters during the dry granulation process and display the dry granulation process status using machine vision;

[0073] The data governance module is used to establish a correspondence between key process parameters and quality attribute parameters in a time series, and to form a multi-dimensional data lake with multiple sets of key process parameters and quality attribute parameters;

[0074] The key process parameter identification module is used to collect relevant data of experimental batches and identify the key process parameters that cause quality fluctuations through a causal discovery method based on Bayesian network learning;

[0075] The process mechanism modeling module uses the CNN algorithm to construct a nonlinear mapping relationship model between key process parameters and quality attribute parameters based on experimental batch related data;

[0076] The parameter optimization module defines the search range of each key process parameter, takes the expected value of the control range of the quality attribute parameter as the optimization target, and uses the optimization algorithm to calculate the optimization space of the key process parameters;

[0077] Human-like intelligent control decision module, used for updating the mapping model based on the optimization interval obtained by the parameter optimization module;

[0078] The control and execution module is used to convert the control variables corresponding to the feasible intervals of key process parameters into control values, input them into the PLC parameter online control system, and execute production process control.

[0079] Furthermore, the multimodal intelligent perception module includes:

[0080] Process quality detection unit, used to obtain quality attribute parameters;

[0081] Process parameter acquisition unit, used to obtain key process parameters;

[0082] The granulation process status calculation visual unit is used to convert the acquired key process parameters into a multi-dimensional state diagram and display the feasible range and real-time values ​​of the key process parameters in real time.

[0083] Furthermore, the process parameter acquisition unit includes: an atomization pressure sensor, an ambient temperature and humidity sensor, a density meter, an infusion frequency acquisition device, and a PLC parameter acquisition subunit. The infusion frequency acquisition device is used to collect the infusion frequency of the extract peristaltic pump, and the PLC parameter acquisition subunit is used to collect the inlet temperature, inlet flow, material temperature, and outlet temperature.

[0084] Furthermore, the dry granulation process parameter control robot system also includes:

[0085] The mobile material collection module is used to obtain particle samples, and the process quality detection unit uses the particle samples obtained by the mobile material collection module to obtain quality attribute parameters.

[0086] Furthermore, the dry granulation process parameter control robot system also includes:

[0087] The remote monitoring module is used to monitor in real time the operating status of the dry granulation process parameter control robot system, key process parameters, quality indicator detection and prediction data, and issue an alarm signal when the quality attribute parameters exceed the control limit; when the key process parameters deviate from the feasible range, issue a process quality warning signal.

[0088] Compared with the prior art, the present invention has the following beneficial effects:

[0089] The AI ​​optimal control method and robotic system for dry granulation process parameters provided by the present invention accurately measures quality attribute parameters, establishes an accurate mapping model between key process parameters and quality attribute parameters, and then obtains the optimized range of key process parameters based on the mapping model and the required quality attribute parameters. Then, according to the real-time key process parameters, the key process parameters can be regulated during the production process. Without measuring the actual quality attribute parameters, the AI-based control system parameter automatic tuning is realized, overcoming the problems of untimely adjustment of traditional production process parameters, low process quality control level, low efficiency, and high dependence on manual experience. At the same time, it also avoids the problems of sample waste and high testing costs caused by frequent sampling.

[0090] The present invention optimizes key process parameters and improves the quality of the pharmaceutical process through AI optimal control method. BRIEF DESCRIPTION OF THE DRAWINGS

[0091] Figure 1 This is a flow chart of the AI ​​optimal control method for dry granulation process parameters provided by the present invention.

[0092] Figure 2 This is a structural block diagram of the AI ​​optimal control robot system for dry granulation process parameters provided by the present invention. DETAILED DESCRIPTION

[0093] The technical solution of the present invention will be clearly described below in conjunction with the accompanying drawings. Obviously, the described embodiments are not all embodiments of the present invention. All other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0094] It should be noted that, unless otherwise specifically stated, the relative arrangements of components and steps, and numerical expressions set forth in these embodiments should not be construed as limiting the scope of the present invention.

[0095] The following description of exemplary embodiments is merely illustrative and is not intended to limit the present invention, its application, or use in any sense. Technologies, methods, and apparatus known to those skilled in the art may not be discussed in detail herein, but to the extent applicable, such technologies, methods, and apparatuses should be considered part of this specification.

[0096] Example 1

[0097] This embodiment provides a method for optimizing the AI ​​control of dry granulation process parameters, such as Figure 1 Shown, including:

[0098] S1. Parameter identification: Identify key process parameters during dry granulation;

[0099] Process parameters in the dry granulation process, such as atomization pressure and ambient temperature and humidity, may affect the final quality attribute parameters, but the degree of influence varies. If every process parameter is included in the construction of the mapping model, the data volume may be too large, which will affect the final model accuracy. Therefore, it is necessary to screen the key process parameters and identify the key process parameters that have a significant impact on the quality attribute parameters.

[0100] In this embodiment, the key process parameters can be determined using a causal relationship algorithm or directly determined based on manual experience. Ultimately, the key process parameters determined in this embodiment are: atomization pressure, ambient temperature and humidity, relative density of incoming material, inlet air temperature, inlet air flow rate, extract peristaltic pump infusion frequency, material temperature, and outlet air temperature.

[0101] The quality attribute parameters in this embodiment include particle size distribution and particle moisture.

[0102] S2. Parameter acquisition: Acquire key process parameters and quality attribute parameters during the dry granulation process. Establish a correspondence between key process parameters and quality attribute parameters in a time series manner, and combine multiple sets of key process parameters and quality attribute parameters into a multi-dimensional data lake.

[0103] Key process parameters and quality attribute parameters are synchronized, outliers are corrected, and missing values ​​are filled; then all data are aligned and fused to obtain a multi-dimensional data lake.

[0104] Different key process parameters are collected using different sensors, and different sensors have different collection frequencies, formats, and qualities. Therefore, the data and quality attribute parameters collected from different sensors are formatted and aligned.

[0105] S3. Constructing a mapping model: Using convolutional neural network modeling, a mapping model between key process parameters and quality attribute parameters is established. Specific methods include:

[0106] S301, input data representation:

[0107] The key process parameters of the input are expressed as a vector X:

[0108] X=[X i ]

[0109] Among them, X i represents the i-th key process parameter;

[0110] Specifically:

[0111] X=[P spray , T env , H env ,ρ input_matter , T inlet , Q inlet , f pump , T material , T outlet ]

[0112] Among them, P spray Indicates the atomization pressure, T env Indicates the ambient temperature, H env represents the ambient humidity, ρ input_matter Indicates the relative density of the incoming material, T inlet Indicates the inlet air temperature, Q inlet Indicates the air flow rate, f pump Indicates the frequency of peristaltic pump infusion of extract, T material Indicates the material temperature, T outlet Indicates the outlet air temperature. Each key process parameter can be a scalar or a series of values ​​with a time series dimension.

[0113] S302, convolutional layer output:

[0114] After the convolution operation, the output feature map F out , feature map F outIt contains multiple extracted features that can characterize the relationship between key process parameters and quality attribute parameters:

[0115]

[0116] Among them, X in (m, n) represents the key process parameters of the input, K(m, n) is the convolution kernel, f is the activation function; the activation function f is the ReLU function; (m, n) represents the spatial coordinates of the feature map;

[0117] S303, output mapping relationship:

[0118] The feature map F is transformed into out Mapped to quality attribute parameters; the output quality attribute parameter formula is:

[0119] [D P , H moisture ]=W·F out +b

[0120] Among them, D P represents the particle size distribution, H moisture represents the particle moisture; W represents the weight matrix, and b is the bias term.

[0121] S4. Parameter optimization: Based on the mapping model, a non-dominated sorting genetic algorithm is used to calculate the optimal intervals of key process parameters corresponding to the required quality attribute parameters; specifically, the following steps are included:

[0122] S401, initialization:

[0123] Randomly generate parent population P t , each individual is a set of key process parameter vectors;

[0124] S402, fast non-dominated sorting and congestion calculation:

[0125] P t Perform fast non-dominated sorting to obtain several non-dominated frontiers F1, F2, ...;

[0126] The individual crowding distance is calculated within each frontier to maintain diversity;

[0127] S403, Selection, Crossover and Mutation:

[0128] The sire pair was selected using the tournament selection method;

[0129] Perform crossover operator to generate offspring Q t ;

[0130] Apply the mutation operator to the offspring;

[0131] S404, Merge and New Generation:

[0132] Merge R t =P t ∪Q t ;

[0133] R t Perform non-dominated sorting again and select the first M individuals to form P according to the frontier and congestion degree. t+1 ;

[0134] S405, termination and output:

[0135] When the maximum algebra G is reached max Or when the population converges, the Pareto optimal solution set composed of all individuals in the first non-dominated frontier F1 and its corresponding parameter interval are output.

[0136] S5. Control decisions and actions: Obtain real-time key process parameters, and based on the optimization range of the key process parameters obtained in S4, determine whether each of the key process parameters is within the corresponding optimization range. If not, adjust the key process parameters until each of the key process parameters is within the corresponding optimization range.

[0137] S6. Model update: establishing a corresponding relationship between the key process parameters adjusted in S5 and the adjusted real-time quality attribute parameters, and updating them into the mapping model.

[0138] The quality attribute parameters of the dry granulation process are obtained in step S2 by:

[0139] A near-infrared spectral camera is used to obtain a spectral image of the particle sample, and a white light camera is used to obtain a white light image of the particle sample. The particle size distribution and particle moisture content of the particle sample are predicted based on the spectral image and the white light image. Specifically, the following steps are involved:

[0140] Predict particle size distribution and particle moisture based on spectral images,

[0141] The specific formula for particle size distribution prediction is:

[0142] D 1P =f NIR (S NIR )

[0143] Among them, D 1P is the particle size distribution predicted by the spectral image, S NIR is the spectral feature of the near-infrared spectrum image, f NIR is the function of the regression model to predict the particle size distribution based on the spectral characteristics;

[0144] The specific formula for particle moisture prediction is:

[0145] H moisture =g NIR (S NIR )

[0146] Among them, H moisture is the particle moisture, S NIR is the spectral feature of the near-infrared spectral image, g NIR is the function of the regression model to predict the particle moisture based on the spectral characteristics;

[0147] The particle size distribution is predicted based on white light images. The specific formula is:

[0148] D 2P =h WL (i WL )

[0149] Among them, D 2P is the particle size distribution predicted by white light imaging, I WL is the geometric feature of the white light image, h WL It is the function of the regression model to predict the particle size distribution based on the geometric characteristics of the white light image;

[0150] The final particle size distribution is:

[0151] D P =α·D 1P +β·D 2P

[0152] Among them, α and β are weight coefficients, α+β=1.

[0153] Use the gradient descent method to determine the weight coefficients α and β and define the objective function

[0154]

[0155] Where N represents the total sample size, represents the true particle size distribution of the i-th sample, D 1P (i) represents the particle size distribution of the i-th sample predicted by the spectral image, D 2P (i) represents the particle size distribution of the i-th sample predicted by the white light image;

[0156] Use the gradient descent method to minimize J(α, β) and obtain the weight coefficients α and β.

[0157] The update rule of the gradient descent method is:

[0158]

[0159] Among them, α' and β' are the updated weight coefficients, and η is the learning rate, which is used to control the size of the step.

[0160] Example 2

[0161] This embodiment provides an AI optimal control robot system for dry granulation process parameters, such as Figure 2 As shown, it includes mobile material collection module, multimodal intelligent perception module, data governance module, key process parameter identification module, process mechanism modeling module, parameter optimization module, human-like intelligent control decision module, control and execution module, and remote monitoring module.

[0162] The mobile retrieving module is primarily comprised of two parts: movement and retrieving. The movement part consists of a mobile robot, which uses laser SLAM navigation to autonomously navigate to the pellet mill's sampling port. The retrieving part consists of a hopper mounted on the end of a robotic arm. During sampling, the robot autonomously navigates to the pellet mill's sample outlet, where the robotic arm automatically opens the sampling port and the pellet mill sample automatically drops into the hopper.

[0163] The multimodal intelligent perception module includes a process quality detection unit, a process parameter acquisition unit, and a granulation process status computational vision unit.

[0164] The process quality inspection unit uses a combination of a near-infrared hyperspectral camera and a white-light camera to capture hyperspectral and white-light images of samples during the granulation process using coaxial illumination. This data is used to predict the particle size distribution and moisture profile of the granules for real-time quality monitoring.

[0165] The process parameter acquisition unit includes multiple sensors that collect key process parameters during the granulation process in real time, such as atomization pressure, ambient temperature and humidity, relative density of incoming material, inlet air temperature and flow rate, extract peristaltic pump infusion frequency, material temperature, and outlet air temperature. This sensor data provides raw data support for subsequent data processing and analysis.

[0166] Among them, the process parameter acquisition unit includes: an atomization pressure sensor, an ambient temperature and humidity sensor, an incoming material relative density meter, an infusion frequency acquisition device, and a PLC parameter acquisition subunit. The infusion frequency acquisition device is used to collect the infusion frequency of the extract peristaltic pump, and the PLC parameter acquisition subunit is used to collect the inlet air temperature, inlet air flow, material temperature, and outlet air temperature.

[0167] The granulation process status calculation and visual unit converts key process parameters such as inlet air temperature, inlet air flow, extract peristaltic pump infusion frequency, and atomization pressure into a high-dimensional state diagram, and displays the feasible range and real-time value of the controllable variables in real time. This "imaging" method helps operators monitor the status of the granulation process in real time.

[0168] The data governance module manages multimodal data through data aggregation, preprocessing, integration, and fusion. This module processes cross-modal data from diverse sources (such as sensors and robots), performing operations such as synchronization, outlier correction, and feature alignment to form a multi-granularity data lake. The data in the data lake includes both batch-level and particle-level data, providing a foundation for subsequent model training and optimization.

[0169] The data aggregation phase begins by acquiring raw data streams from various data sources, which typically have varying collection frequencies, formats, and qualities.

[0170] In the data preprocessing stage, the raw data are synchronized, outlier corrected and missing values ​​filled.

[0171] In the data integration and fusion stage, data streams of different modalities are aligned and fused to obtain a unified representation.

[0172] The key process parameter identification module collects relevant data of experimental batches and identifies the key process parameters that cause quality fluctuations through a causal discovery method based on Bayesian network learning, thereby determining the key process parameters in the dry granulation process.

[0173] The process mechanism modeling module uses the convolutional neural network (CNN) algorithm to perform nonlinear mapping of multimodal key process parameters such as inlet air temperature, inlet air flow, material temperature, outlet air temperature, infusion frequency, ambient temperature and humidity based on experimental batch related data, to determine their relationship with quality attribute parameters (particle moisture, particle size distribution) and establish a mapping model.

[0174] The parameter optimization module takes particle moisture and particle size distribution as optimization targets, and uses multi-objective optimization algorithms such as particle swarm optimization, simulated annealing algorithm, or non-dominated sorting genetic algorithm to calculate the optimization range of each key process parameter.

[0175] The human-like intelligent control decision-making module learns the PID curve through full-process (time series) simulation to form a typical pattern curve cluster of key process parameters such as inlet air temperature, inlet air flow, extract peristaltic pump infusion frequency, and atomization pressure (that is, the pattern corresponding to the predicted quality indicators such as particle moisture and particle size distribution).

[0176] After implementation, the system collects actual production batch time series data for key process parameters and quality attribute parameters. Specifically, the mobile material collection module sequentially samples at set time intervals and uses the multimodal intelligent sensing module to detect quality attribute parameters such as sample moisture and particle size. These are used as network outputs. Key process parameters such as inlet temperature and flow rate before the sampling point within the batch are used as network inputs. The system continuously learns, automatically iteratively updates the mapping model, and continuously optimizes the feasible range of key process parameters.

[0177] With the continuous accumulation of production and testing data, a series of process quality pattern curve clusters are obtained through autonomous learning, which can simulate human thinking to make control decisions; specifically, the control decision rules are first obtained through the automatic search for optimal reasoning methods such as the time series similarity measurement algorithm, that is, a control decision model is established, and then the mode to be adopted is decided according to the actual variation of the key process parameters, and then the operating variable values ​​corresponding to the key process parameters are automatically set. Under given input conditions, the parameter control PID curve of the appropriate mode is automatically generated, and feedback control is further realized according to the actual production situation.

[0178] The control and execution module converts the control variables corresponding to the feasible ranges of key process parameters such as atomization pressure, ambient temperature and humidity, relative density of incoming materials, inlet air temperature, inlet air flow, extract peristaltic pump infusion frequency, material temperature, and outlet air temperature into operational control values, inputs them into the PLC parameter online control system, and executes production process control.

[0179] Specifically, the parameter optimization module uses algorithmic optimization to determine optimal key process parameters, such as inlet air temperature, inlet air flow rate, extract peristaltic pump infusion frequency, and atomization pressure. These parameters are crucial for ensuring ideal drug granule quality. In the control and execution module, these optimal key process parameters are input into the PLC in real time for online feedback control. The PLC is responsible for receiving the key process parameters output by the optimization module and converting them into specific control instructions. These control instructions act on the dry granulation process to ensure that key process parameters are always maintained within the optimized range.

[0180] Control instructions are generated and written to the PLC via OPC UA. The feedforward correction is triggered by the density fluctuation of the incoming extract through the feedforward and feedback collaborative mechanism, and feedback iteration is triggered by the real-time hyperspectral deviation, thereby ensuring that the system can adjust parameters in time to avoid quality fluctuations caused by raw material density fluctuations, and correct and optimize the control strategy through real-time feedback.

[0181] The remote monitoring module provides real-time monitoring of process data, including the robot system's operating status, detection data, prediction data, and control parameters. Through a remote communication interface, operators can access this data and conduct real-time monitoring from anywhere. When critical quality attributes such as particle moisture and particle size distribution exceed set control limits, the system automatically issues an alarm. Alternatively, when key process parameters deviate from acceptable ranges, the system automatically issues a process quality warning signal and prompts intervention. This module ensures that any anomalies or deviations in the production process are quickly detected, allowing necessary adjustments to be implemented to ensure efficient and high-quality production.

[0182] The above specific embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to examples, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the scope of the technical solutions of the present invention, and all of these should be included in the scope of the claims of the present invention.

Claims

1. The AI ​​optimal control method for dry granulation process parameters is characterized by: include: S1. Parameter identification: Identify key process parameters during dry granulation; S2. Parameter acquisition: Acquire key process parameters and quality attribute parameters during the dry granulation process. Establish a correspondence between key process parameters and quality attribute parameters in a time series manner, and combine multiple sets of key process parameters and quality attribute parameters into a multi-dimensional data lake. S3. Build a mapping model: Use convolutional neural network modeling to establish a mapping model between key process parameters and quality attribute parameters; S4. Parameter optimization: Calculating the optimization intervals of key process parameters corresponding to the required quality attribute parameters based on the mapping model; S5. Control decisions and actions: Obtain real-time key process parameters, and based on the optimization range of the key process parameters obtained in S4, determine whether each of the key process parameters is within the corresponding optimization range. If not, adjust the key process parameters until each of the key process parameters is within the corresponding optimization range.

2. The dry granulation process parameter AI optimal control method according to claim 1, characterized in that: In step S2, multiple sets of key process parameters and quality attribute parameters are combined into a multi-dimensional data lake, specifically including: Key process parameters and quality attribute parameters are synchronized, outliers are corrected, and missing values ​​are filled; then all data are aligned and fused to obtain a multi-dimensional data lake.

3. The dry granulation process parameter AI optimal control method according to claim 1, characterized in that: In step S3, a convolutional neural network is used to construct a mapping model between key process parameters and quality attribute parameters. The specific method includes: S301, input data representation: The key process parameters of the input are expressed as a vector X: X=[X i ] Among them, X i represents the i-th key process parameter; S302, convolutional layer output: After the convolution operation, the output feature map F out , feature map F out Characterize the relationship between key process parameters and quality attribute parameters: Among them, X in (m, n) represents the key process parameters of the input, K(m, n) is the convolution kernel, and f is the activation function; (m, n) represents the spatial coordinates of the feature map; S303, output mapping relationship: The feature map F is transformed into out Mapped to quality attribute parameters; the output quality attribute parameter formula is: [Y j ]=W·F out +b Among them, Y j represents the j-th quality attribute parameter; W represents the weight matrix, and b is the bias term.

4. The dry granulation process parameter AI optimal control method according to claim 3, characterized in that: Key process parameters include atomization pressure, ambient temperature and humidity, relative density of incoming materials, inlet air temperature, inlet air flow, extract peristaltic pump infusion frequency, material temperature, and outlet air temperature; Quality attribute parameters include particle size distribution and particle moisture.

5. The dry granulation process parameter AI optimal control method according to claim 4, characterized in that: In step S4, based on the mapping model, a non-dominated sorting genetic algorithm is used to calculate the optimization interval of the key process parameters corresponding to the required quality attribute parameters; Specifically include: S401, initialization: Randomly generate parent population P t , each individual is a set of key process parameter vectors; S402, fast non-dominated sorting and congestion calculation: P t Perform fast non-dominated sorting to obtain several non-dominated frontiers F1, F2, ...; The individual crowding distance is calculated within each frontier to maintain diversity; S403, Selection, Crossover and Mutation: The sire pair was selected using the tournament selection method; Perform crossover operator to generate offspring Q t ; Apply the mutation operator to the offspring; S404, Merge and New Generation: Merge R t =P t ∪Q t ; R t Perform non-dominated sorting again and select the first M individuals to form P according to the frontier and congestion degree. t+1 ; S405, termination and output: When the maximum algebra G is reached nax Or when the population converges, the Pareto optimal solution set composed of all individuals in the first non-dominated frontier F1 and its corresponding parameter interval are output.

6. The dry granulation process parameter AI optimal control method according to claim 1, characterized in that: The AI ​​optimal control method for dry granulation process parameters also includes: S6. Model update: establishing a corresponding relationship between the key process parameters adjusted in S5 and the adjusted real-time quality attribute parameters, and updating them into the mapping model.

7. The dry granulation process parameter AI optimal control method according to claim 4, characterized in that: In step S2, the quality attribute parameters of the dry granulation process are obtained by: A near-infrared spectral camera is used to obtain a spectral image of the particle sample, and a white light camera is used to obtain a white light image of the particle sample. The particle size distribution and particle moisture of the particle sample are predicted based on the spectral image and the white light image.

8. The method for optimal control of dry granulation process parameter AI according to claim 7, characterized in that: Predict the particle size distribution and moisture content of particle samples based on spectral images and white light images, including: Predict particle size distribution and particle moisture based on spectral images, The specific formula for particle size distribution prediction is: D 1P =f NIR (S NIR ) Among them, D 1P is the particle size distribution predicted by the spectral image, S NIR is the spectral feature of the near-infrared spectrum image, f NIR is the function of the regression model to predict the particle size distribution based on the spectral characteristics; The specific formula for particle moisture prediction is: H moisture =g NIR (S NIR ) Among them, H moisture is the particle moisture, S NIR is the spectral feature of the near-infrared spectrum image, g NIR is the function of the regression model to predict the particle moisture based on the spectral characteristics; The particle size distribution is predicted based on white light images. The specific formula is: D 2P =h WL (I WL ) Among them, D 2P is the particle size distribution predicted by white light imaging, I WL is the geometric feature of the white light image, h WL It is the function of the regression model to predict the particle size distribution based on the geometric characteristics of the white light image; The final particle size distribution is: D P =α·D 1P +β·D 2P Among them, α and β are weight coefficients, α+β=1.

9. The method for optimal control of dry granulation process parameter AI according to claim 8, characterized in that: Use the gradient descent method to determine the weight coefficients α and β and define the objective function Where N represents the total sample size, represents the true particle size distribution of the i-th sample, D 1P (i) represents the particle size distribution of the i-th sample predicted by the spectral image, D 2P (i) represents the particle size distribution of the i-th sample predicted by the white light image; Use the gradient descent method to minimize J(α, β) and obtain the weight coefficients α and β.

10. The dry granulation process parameter AI optimal control method according to claim 9, characterized in that: The update rule of the gradient descent method is: Among them, α' and β' are the updated weight coefficients, and η is the learning rate.

11. AI-based optimal control robot system for dry granulation process parameters, including: A multimodal intelligent perception module is used to obtain key process parameters and quality attribute parameters during the dry granulation process and display the dry granulation process status using machine vision; The data governance module is used to establish a correspondence between key process parameters and quality attribute parameters in a time series, and to form a multi-dimensional data lake with multiple sets of key process parameters and quality attribute parameters; The key process parameter identification module is used to collect relevant data of experimental batches and identify the key process parameters that cause quality fluctuations through a causal discovery method based on Bayesian network learning; The process mechanism modeling module uses the CNN algorithm to construct a nonlinear mapping relationship model between key process parameters and quality attribute parameters based on experimental batch related data; The parameter optimization module defines the search range of each key process parameter, takes the expected value of the control range of the quality attribute parameter as the optimization target, and uses the optimization algorithm to calculate the optimization space of the key process parameters; Human-like intelligent control decision module, used for updating the mapping model based on the optimization interval obtained by the parameter optimization module; The control and execution module is used to convert the control variables corresponding to the feasible intervals of key process parameters into control values, input them into the PLC parameter online control system, and execute production process control.

12. The dry granulation process parameter AI optimal control robot system according to claim 11, characterized in that: The multimodal intelligent perception module includes: Process quality detection unit, used to obtain quality attribute parameters; Process parameter acquisition unit, used to obtain key process parameters; The granulation process status calculation visual unit is used to convert the acquired key process parameters into a multi-dimensional state diagram and display the feasible range and real-time values ​​of the key process parameters in real time.

13. The dry granulation process parameter AI optimal control robot system according to claim 12, characterized in that: The dry granulation process parameter control robot system also includes: The mobile material collection module is used to obtain particle samples, and the process quality detection unit uses the particle samples obtained by the mobile material collection module to obtain quality attribute parameters.

14. The dry granulation process parameter AI optimal control robot system according to claim 11, characterized in that: The dry granulation process parameter control robot system also includes: The remote monitoring module is used to monitor in real time the operating status of the dry granulation process parameter control robot system, key process parameters, quality indicator detection and prediction data, and issue an alarm signal when the quality attribute parameters exceed the control limit; when the key process parameters deviate from the feasible range, issue a process quality warning signal.

Images