Drug jet milling self-adaptive modeling control method and system based on particle size distribution

Abstract

The invention discloses a drug jet milling self-adaptive modeling control method and system based on particle size distribution, and relates to the technical field of self-adaptive control, and the method comprises the steps: building a mixed digital twinborn model; connecting a particle size monitoring network, establishing particle size distribution data according to the collection and monitoring distribution relationship, and obtaining process parameters; process parameters are input into the mechanism model for crushing state simulation, model parameter iteration updating is carried out through the data driving model based on particle size distribution data, and parameter updating is carried out on the mixed digital twinborn model in an interactive mode; carrying out difference analysis, and determining a crushing compensation target particle size; and a reverse optimization model is established, self-adaptive search of reverse crushing parameters is carried out according to the crushing compensation target particle size, crushing control parameters are obtained, and the crushing control parameters are fed back to a controller for crushing airflow self-adaptive control. The technical problems that in the prior art, the particle size deviation is large and the stability is poor in the medicine airflow smashing process are solved, and the technical effect of achieving real-time accurate control over particle size distribution is achieved.

Description

Technical Field

[0001] This invention relates to the field of adaptive control technology, specifically to an adaptive modeling control method and system for pharmaceutical airflow pulverization based on particle size distribution. Background Technology

[0002] In the process of pharmaceutical air jet milling, particle crushing kinetics and classification kinetics are coupled, and process conditions such as feed rate, classifier speed, and milling air pressure are all in a dynamic state, making it difficult to maintain a stable particle size distribution. Due to the lack of precise modeling and effective control of particle size distribution, large particle size deviations and uneven distribution are prone to occur during the milling process. Summary of the Invention

[0003] This application provides an adaptive modeling control method and system for pharmaceutical airflow pulverization based on particle size distribution, which is used to address the technical problems of large particle size deviation and poor stability in the pharmaceutical airflow pulverization process in the prior art.

[0004] In view of the above problems, this application provides an adaptive modeling control method and system for drug airflow pulverization based on particle size distribution.

[0005] The first aspect of this application provides an adaptive modeling and control method for pharmaceutical airflow pulverization based on particle size distribution, the method comprising: A hybrid digital twin model is established, including a mechanistic model and a data-driven model. A particle size monitoring network is connected, and particle size distribution data is established according to the collected monitoring distribution relationship. Process parameters are acquired, including at least the stager speed, feed rate, and pulverizing air pressure. These process parameters are input into the mechanistic model to simulate the pulverizing state. Based on the particle size distribution data, the model parameters are iteratively updated using the data-driven model, and the hybrid digital twin model is interactively updated with updated parameters. Based on the real-time monitored particle size distribution data and the target particle size distribution, differential analysis is performed to determine the pulverizing compensation target particle size. Based on the pulverizing compensation target particle size and the updated hybrid digital twin model, a reverse optimization model is established. Adaptive search of reverse pulverizing parameters is performed according to the pulverizing compensation target particle size to obtain pulverizing control parameters, which are then fed back to the controller for adaptive pulverizing airflow control.

[0006] A second aspect of this application provides an adaptive modeling and control system for pharmaceutical airflow pulverization based on particle size distribution, the system comprising: The model building module is used to build a hybrid digital twin model, including a mechanistic model and a data-driven model; the monitoring module is used to connect to the particle size monitoring network, build particle size distribution data according to the collected monitoring distribution relationship, and obtain process parameters, including at least the classifier wheel speed, feed rate, and pulverizing air pressure; the parameter update module is used to input the process parameters into the mechanistic model to simulate the pulverizing state, and to iteratively update the model parameters based on the particle size distribution data through the data-driven model, interactively updating the parameters of the hybrid digital twin model; the differential analysis module is used to perform differential analysis based on the real-time monitored particle size distribution data and the target particle size distribution to determine the pulverizing compensation target particle size; the adaptive control module is used to build a reverse optimization model based on the pulverizing compensation target particle size and the updated hybrid digital twin model, to perform an adaptive search of reverse pulverizing parameters according to the pulverizing compensation target particle size, to obtain pulverizing control parameters, and to feed them back to the controller for adaptive control of the pulverizing airflow.

[0007] One or more technical solutions provided in this application have at least the following technical effects or advantages: This application establishes a hybrid digital twin model, including a mechanistic model and a data-driven model; connects to a particle size monitoring network, establishes particle size distribution data according to the collected monitoring distribution relationship, and obtains process parameters, including at least the classifier wheel speed, feed rate, and pulverizing air pressure; inputs the process parameters into the mechanistic model to simulate the pulverizing state, and iteratively updates the model parameters through the data-driven model based on the particle size distribution data, interactively updating the parameters of the hybrid digital twin model; performs differential analysis based on the real-time monitored particle size distribution data and the target particle size distribution to determine the pulverizing compensation target particle size; establishes a reverse optimization model based on the pulverizing compensation target particle size and the updated hybrid digital twin model, performs adaptive search of reverse pulverizing parameters according to the pulverizing compensation target particle size, obtains pulverizing control parameters, and feeds them back to the controller for adaptive control of the pulverizing airflow. This invention addresses the technical problems of large particle size deviation and poor stability in the process of air jet milling of pharmaceuticals in existing technologies. By establishing a hybrid digital twin model that combines a mechanistic model and a data-driven model, and performing differential analysis and reverse optimization parameter search based on real-time particle size distribution data, this invention achieves real-time, effective, and precise control of pharmaceutical air jet milling using an intelligent control system. It also considers the specific environmental characteristics of particle size distribution, thereby improving the adaptability of milling control and the robustness of the system. Attached Figure Description

[0008] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 A schematic diagram of the adaptive modeling and control method for drug airflow pulverization based on particle size distribution provided in this application embodiment; Figure 2 A schematic diagram of the structure of the adaptive modeling and control system for drug airflow pulverization based on particle size distribution provided in the embodiments of this application.

[0010] Figure labeling: Model building module 11, monitoring module 12, parameter update module 13, differential analysis module 14, adaptive control module 15. Detailed Implementation

[0011] This application provides an adaptive modeling and control method and system for pharmaceutical airflow pulverization based on particle size distribution. It addresses the technical problems of large particle size deviation and poor stability in existing pharmaceutical airflow pulverization processes. By establishing a hybrid digital twin model combining a mechanistic model and a data-driven model, and performing differential analysis and inverse optimization parameter search based on real-time particle size distribution data, it achieves real-time, effective, and precise control during pharmaceutical airflow pulverization using an intelligent control system. This approach considers the specific environmental characteristics of particle size distribution, improving the adaptability and robustness of the pulverization control.

[0012] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0013] It should be noted that any variation of the terms "comprising" and "having" is intended to cover non-exclusive inclusion, for example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to such processes, methods, products, or devices.

[0014] Example 1, as Figure 1 As shown, this application provides an adaptive modeling and control method for drug airflow pulverization based on particle size distribution, the method comprising: Step S100: Establish a hybrid digital twin model, including a mechanistic model and a data-driven model.

[0015] In this embodiment, a mechanistic model characterizing crushing and classification dynamics is first constructed based on the crusher's operating mechanism and sample experimental data. Then, a data-driven model is trained within a neural network framework, and dynamic correction is achieved by adjusting the mechanistic model parameters through real-time process data input and output. Finally, a digital twin model is built based on the crusher's physical structure, and the data-driven model, as an internal learning module, forms an interactive learning path with the mechanistic model. The output is dynamically injected as real-time updated parameters, thus completing the establishment of a hybrid digital twin model.

[0016] Furthermore, the method provided in the application embodiments, which establishes a hybrid digital twin model, including a mechanistic model and a data-driven model, also includes: Based on the operating mechanism of the pulverizer and sample experimental data, a mechanistic model is constructed to characterize the crushing and classification dynamics of particles during airflow pulverization. A data-driven model is trained based on a neural network framework to perform real-time correction of the mechanistic model. Its input data is real-time process data, and its output is the compensation values ​​for the parameters in the mechanistic model. The physical structure of the pulverizer is collected and combined with the mechanistic model to establish a digital twin model. The data-driven model is used as an internal learning module to establish an interactive learning path with the mechanistic model, and the hybrid digital twin model is obtained by integration. The integration method is to dynamically inject the output of the data-driven module as real-time updated parameters into the mechanistic model.

[0017] Furthermore, the method provided in the application embodiments also includes: The mechanism model includes a crushing rate function and a grading efficiency function. The data-driven model is a data-driven correction model. The model output compensation value is the dynamic compensation amount of the key function parameters in the mechanism model, including the crushing rate compensation amount and the grading efficiency compensation amount.

[0018] In this embodiment, a mechanistic model is first established based on the operating mechanism of the pulverizer and sample experimental data. This mechanistic model characterizes the particle crushing and classification dynamics during airflow pulverization, with its core being the establishment of a crushing rate function and a classification efficiency function. The crushing rate function is obtained by analyzing the energy absorption law of particles under airflow and fitting it with experimental crushing data of particles under different pressure and flow rates, describing the rate law of particle transformation from large to small diameter per unit time. The classification efficiency function is obtained by modeling the correlation between the speed of the classifier wheel and the particle trajectory, combined with actual classification experimental data, characterizing the probability distribution of particles of different diameters passing through or being trapped in the classification stage. Through the above methods, the mechanistic model achieves a mathematical description of the crushing and classification processes.

[0019] Based on the established mechanistic model, a data-driven correction model is trained using a neural network framework. This model takes real-time process data as input, including at least the stager wheel speed, feed rate, and crushing air pressure. The output is compensation values ​​for key function parameters in the mechanistic model. These compensation values ​​include crushing rate compensation and staged efficiency compensation, used to dynamically correct the crushing rate function and staged efficiency function in the mechanistic model. By introducing the nonlinear fitting and self-learning capabilities of the neural network, the data-driven correction model automatically outputs compensation values ​​and corrects the mechanistic model when discrepancies exist between the real-time particle size distribution monitoring data and the mechanistic model's predictions, thus ensuring that the predicted particle size distribution gradually converges with the actual particle size distribution.

[0020] After constructing the mechanistic model and the data-driven correction model, a digital twin model is established by combining the physical structural information of the crusher. The digital twin model is a virtual mapping of the crusher and its operating process, achieving operational simulation consistent with the actual equipment conditions. The digital twin model inherits the physical mechanism description of the mechanistic model and provides a virtual environment for the real-time compensation of the data-driven correction model, enabling the two types of models to run collaboratively on the same virtual platform, ensuring the dynamic correspondence between virtual predictions and the actual process.

[0021] Building upon the digital twin model, a data-driven correction model is embedded as an intrinsic learning module, establishing an interactive learning pathway with the mechanistic model. This interactive learning pathway dynamically injects the breakage rate compensation and grading efficiency compensation output from the data-driven correction model as real-time update parameters into the mechanistic model, adaptively correcting the breakage rate function and grading efficiency function. Through this approach, the physical interpretability of the mechanistic model and the real-time correction capability of the data-driven correction model are integrated, ultimately resulting in a hybrid digital twin model.

[0022] Step S200: Connect the particle size monitoring network, establish particle size distribution data according to the collected monitoring distribution relationship, and obtain process parameters, including at least the classifier wheel speed, feed rate, and crushing air pressure.

[0023] In this embodiment, a particle size monitoring network is used to collect particle size distribution data in real time. The particle size monitoring network employs laser diffraction as the detection method to monitor the distribution relationship of particles in different particle size ranges and generates particle size distribution data based on the collected results. This particle size distribution data, in the form of distribution curves, characterizes the proportional characteristics of particles within each particle size range, reflecting the particle size characteristics of the pulverized product. Simultaneously with the collection of particle size distribution data, the particle size monitoring network acquires process parameters, including the classifier rotation speed, feed rate, and pulverizing air pressure. The classifier rotation speed characterizes the classification efficiency, the feed rate characterizes the material supply intensity, and the pulverizing air pressure characterizes the airflow energy level.

[0024] Step S300: Input the process parameters into the mechanism model to simulate the crushing state, and perform iterative updates of the model parameters through data-driven model based on the particle size distribution data, and interactively update the parameters of the hybrid digital twin model.

[0025] In this embodiment, the real-time acquired process parameters are first input into the mechanistic model. The crushing state is simulated using the model's internal crushing rate function and classification efficiency function, outputting the predicted particle size distribution under the corresponding working conditions. Then, this predicted particle size distribution is compared with the real-time monitored particle size distribution data to obtain the particle size distribution difference. Based on this difference, a recursive estimation algorithm is used to iteratively update the model parameters of the data-driven model online, gradually reducing the output compensation value by decreasing the particle size distribution difference until a preset error threshold is met. The finally determined optimal model parameters are used to update the parameters of the mechanistic model in the hybrid digital twin model.

[0026] Furthermore, in the method provided in the application embodiments, the process parameters are input into the mechanism model to simulate the crushing state, and the model parameters are iteratively updated through data-driven modeling based on the particle size distribution data. The method also includes interactively updating the parameters of the hybrid digital twin model. The real-time collected process parameters are used as input to the mechanism model to simulate the state of the crushing process, and the predicted particle size distribution under the current working conditions is output. The predicted particle size distribution is compared with the real-time collected particle size distribution data to obtain the particle size distribution difference. Based on the particle size distribution difference, the model parameters of the data-driven model are updated online using a recursive estimation algorithm, so that the output compensation value of the data-driven model can reduce the particle size distribution difference. The search is iterated until the error threshold is met, and the optimal model parameters are determined to update the parameters of the mechanism model in the hybrid digital twin model.

[0027] In this embodiment, the real-time acquired process parameters are first input into the mechanism model, and the state simulation of the crushing process is carried out based on the crushing rate function and the classification efficiency function. Discrete-time step simulation is used to iteratively solve the crushing and classification process of each particle size range, and output the predicted particle size distribution under the current working condition.

[0028] Next, the predicted particle size distribution is compared with the real-time collected particle size distribution data. The difference is calculated point by point according to the particle size segment, and the weighted mean square error is summarized to obtain the particle size distribution difference used to quantify the model bias.

[0029] Subsequently, based on the difference in particle size distribution, a recursive estimation algorithm is used to iteratively update the model parameters of the data-driven model online, so that the compensation values ​​output by the data-driven model, including the crushing rate compensation and the classification efficiency compensation, gradually reduce the difference in particle size distribution. In this process, a preset error threshold is used as the convergence criterion, and the optimal model parameters are determined when the threshold is met. These parameters are then injected into the mechanism model in the hybrid digital twin model to complete the parameter update, realizing interactive correction driven by mechanism and data.

[0030] Step S400: Based on the real-time monitored particle size distribution data, perform differential analysis in conjunction with the target particle size distribution to determine the target particle size for crushing compensation.

[0031] In this embodiment, the particle size distribution data obtained from real-time monitoring is first compared with the preset target particle size distribution within the same particle size range. The difference between the two is calculated segment by segment to obtain a difference curve reflecting the difference between the predicted state and the target requirement. When the difference curve shows a positive value in a certain range, it indicates that there are more coarse particles in that particle size range. When a negative value appears, it indicates that there are more fine particles in that particle size range. Combining the extreme points of the difference curve, the most significant deviation location is identified, and the particle size corresponding to that location is determined as the target particle size for crushing compensation.

[0032] Step S500: Based on the target particle size for crushing compensation and the updated hybrid digital twin model, establish a reverse optimization model, perform an adaptive search for reverse crushing parameters according to the target particle size for crushing compensation, obtain crushing control parameters, and feed them back to the controller for adaptive control of the crushing airflow.

[0033] In this embodiment, an inverse optimization model is first established based on the target particle size for crushing compensation and an updated hybrid digital twin model. The target particle size for crushing compensation is then used as the control target for inverse simulation. Based on the mechanistic model, inverse simulation is performed to obtain a sequence of inverse control parameters, which includes key process parameters such as the stager speed, feed rate, and crushing air pressure. Subsequently, the inverse control parameter sequence is reversed in the forward sequence, and the parameters are verified using real-time collected particle size distribution data as the initial state and the target particle size distribution as the final target, evaluating the consistency between the simulation final state and the target distribution. When the verification results meet the preset accuracy requirements, the crushing control parameters are determined. These crushing control parameters are fed back to the controller, which executes corresponding adjustment operations to achieve adaptive control of the crushing airflow, ensuring that the particle size distribution converges to the target particle size distribution and remains stable.

[0034] Furthermore, in the method provided in the application embodiments, based on the target particle size for crushing compensation and the updated hybrid digital twin model, an inverse optimization model is established, and an adaptive search of inverse crushing parameters is performed according to the target particle size for crushing compensation to obtain crushing control parameters. This further includes: Based on the updated hybrid digital twin model, a reverse optimization model is established with the crushing compensation target particle size as the control target. Based on the mechanism model, reverse simulation is performed to obtain the reverse control parameter sequence. According to the reverse control parameter sequence, the forward sequence is reversed, and the parameters are verified with the collected particle size distribution data as the initial step and the target particle size distribution as the final target. When the parameter verification requirements are met, the crushing control parameters are determined.

[0035] In this embodiment, a reverse optimization model is first established based on the updated hybrid digital twin model, with the target particle size for crushing compensation as the objective. During model construction, a control timeline is built using the crushing process and processing time as constraints. Based on the mechanistic model, the control timeline is segmented into stages to determine the target particle size for each node constraint and to complete node labeling. Subsequently, based on the control timeline and the target particle size for each node constraint, a probabilistic search algorithm is used for reverse deduction. Candidate parameter sequences are evaluated through multiple forward simulation iterations, and the optimal parameter sequence is selected and confirmed as the reverse control parameter sequence.

[0036] After obtaining the inverse control parameter sequence, its forward sequence is reversed. Using the collected particle size distribution data as the initial state, the forward-arranged control parameter sequence is input into the updated hybrid digital twin model for forward simulation. The matching degree between the particle size distribution of the final simulation state and the target particle size distribution is obtained. When the matching degree reaches a preset threshold, the control parameter sequence is confirmed to be valid, the parameter verification is completed, and the control parameter sequence is finally determined as the crushing control parameter.

[0037] Furthermore, in the method provided in the application embodiments, based on the updated hybrid digital twin model, using the crushing compensation target particle size as the control target, an inverse optimization model is established, and inverse simulation deduction is performed based on the mechanism model to obtain the inverse control parameter sequence, which further includes: A control time sequence chain is constructed with the crushing process and processing time as constraints and the crushing compensation target particle size as the objective. Based on the mechanistic model, the control time sequence chain is divided into stages with constraints, the node constraint target particle size is determined, and the nodes are labeled. Based on the control time sequence chain and the node constraint target particle size, a probabilistic search algorithm is used for reverse deduction. Candidate parameter sequences are evaluated through multiple forward simulation iterations, and the optimal control parameter sequence is selected based on the deduction results.

[0038] In this embodiment, when constructing a control timing chain with the crushing process and processing time as constraints and the target particle size for crushing compensation as the objective, the starting and ending points of the processing cycle are first determined based on the actual running time of the crushing process, and these are used as boundary conditions for the time axis. Then, a fixed-step discretization method is used to divide the processing time, equally dividing the entire processing cycle into several time nodes, each corresponding to a specific moment in the crushing process. Next, combined with the target particle size for crushing compensation, a function fitting method is used to generate a reference trajectory that gradually converges from the initial particle size distribution to the target particle size distribution, and sampling is performed at the aforementioned time nodes to obtain the target particle size value corresponding to each node. Finally, all nodes are arranged in chronological order to form the control timing chain.

[0039] Next, the control time-series chain is segmented into stages based on the mechanistic model. In this process, firstly, the dominant time constant of the process is extracted by analyzing the response characteristics of crushing and staged dynamics. Based on this dominant time constant, the prediction step size and prediction time domain of the inverse optimization model are determined, thereby calculating the number of intermediate nodes. Subsequently, with the crushing compensation target particle size as the objective, a desired reference trajectory is generated by combining the dynamic response characteristics of the mechanistic model. This reference trajectory is then discretely sampled according to the set number of time nodes to obtain the node constraint target particle size corresponding to each stage. Based on this, each node is labeled to form a time-series constraint.

[0040] Subsequently, based on the control timeline chain and the target particle size of the node constraints, a probabilistic search algorithm is used for reverse inference. During the inference process, a probabilistic graph search tree is constructed with the current state as the root node and the final state of the timeline chain as the target. The upper confidence interval algorithm is used to recursively select child nodes in the search tree, continuously expanding new candidate control actions and forming multiple feasible paths. Then, through multiple forward simulation iterations, a hybrid digital twin model is invoked to generate a complete state prediction trajectory, and a reward value is calculated for each trajectory. This reward value comprehensively evaluates the degree of achievement of the final state target and the degree of satisfaction of the target particle size of the intermediate node constraints. Finally, based on the inference results, the path with the highest reward value is selected from the candidate parameter sequence, and its corresponding control action sequence is extracted and determined as the optimal control parameter sequence. This optimal control parameter sequence is the reverse control parameter sequence.

[0041] Furthermore, the method provided in the application embodiments, which employs a probabilistic search algorithm for reverse deduction, further includes: Using the current state as the root node and the final state of the time-series chain as the target, a probabilistic graph search tree is initialized. Nodes in the search tree represent states at specific times, and edges represent candidate control actions connecting states. Starting from the root node, the optimal child node is recursively selected based on the upper confidence interval algorithm until a scalable non-final state node is reached. One or more new child nodes are added to the selected node, each corresponding to a candidate control action that has not been fully explored. Starting from the newly expanded node, simulation is performed based on the hybrid digital twin model to the final state of the time-series chain, resulting in a complete state prediction trajectory. The reward value of the prediction trajectory is calculated, and the reward value comprehensively evaluates the degree of achievement of the final state target and the degree of constraint violation of intermediate nodes. The statistical information of all visited nodes is updated in reverse along the search path. After a predetermined number of iterations, the path with the highest comprehensive reward value is selected from the search tree, and the control action sequence corresponding to the path is extracted to obtain the inverse control parameter sequence.

[0042] In this embodiment, the current state is used as the root node, and the final state of the time-series chain is used as the target to initialize a probabilistic graphical search tree. In this search tree, nodes represent the operating state at a specific moment, and edges represent candidate control parameters that may be taken between these states, such as the stager speed, feed rate, or pulverizing air pressure. In this way, a correspondence between states and control parameters is established.

[0043] During the search process, starting from the root node, a recursive search is performed using the upper confidence interval algorithm. At each recursive step, the upper confidence interval algorithm determines the priority of child nodes based on the number of visits to the node and its statistical upper bound, thus deciding the next direction of exploration, until a scalable node that has not yet reached its final state is reached.

[0044] At the reached scalable node, a node expansion operation is performed to add new child nodes to that node. Each child node represents an unexplored candidate control parameter. For example, when the current node corresponds to a feed rate of 10 kg / h, the newly added child node may correspond to a feed rate of 12 kg / h or a pulverizing air pressure of 0.65 MPa, thereby expanding new possible paths.

[0045] After the expansion is complete, starting from the newly expanded node, a hybrid digital twin model is invoked for forward simulation, tracing along the time axis to the final state of the time-series chain, generating a complete state prediction trajectory. This trajectory records the evolution of the particle size distribution over time under the selected control parameters, such as the predicted particle size distribution results at the 5th, 10th, and 20th seconds.

[0046] Once the predicted trajectory is obtained, the reward value calculation begins. The reward value consists of two parts. The first part is the final state target achievement rate, assessed by the deviation between the predicted particle size distribution and the target particle size distribution, often using mean squared error (MSE) as a metric. For example, when the MSE between the predicted and target particle size distributions is 0.02, it indicates that the final state is relatively close to the target. The second part is the degree of constraint violation at intermediate nodes. Error checks are performed at each node at the constraint target particle size. If the deviation between the predicted value and the constraint target particle size at a node exceeds a 5% threshold, the reward value is reduced according to the magnitude of the deviation. For example, if two out of ten nodes violate the constraint, and each violation deducts 0.01, a total of 0.02 is deducted from the original reward. The final reward value is the weighted sum of the two parts, used to quantify the overall quality of the trajectory.

[0047] Once the reward value is determined, a reverse update is performed along the path that generated the trajectory in the search tree. The updates include increasing the number of node visits and correcting node statistics based on the new reward value, allowing the search tree to gradually accumulate experience. After multiple iterations, a large number of candidate paths will be formed in the search tree.

[0048] After a preset number of iterations, the path with the highest comprehensive reward value is selected from the search tree, and the combination of control parameters on that path is extracted, such as a classifier wheel speed of 8000 rpm, a feed rate of 12 kg / h, and a crushing air pressure of 0.6 MPa. This parameter combination is the inverse control parameter sequence.

[0049] Furthermore, in the method provided in the application embodiments, the stage constraint segmentation of the control timing chain based on the mechanistic model to determine the target particle size of the node constraints further includes: Based on the aforementioned mechanism model, the response characteristics of crushing dynamics and classification dynamics are analyzed, and the dominant time constant of the process is extracted. Based on the dominant time constant, the prediction step size and prediction time domain of the inverse optimization model are determined, and the number of intermediate nodes is calculated. Taking the crushing compensation target particle size as the objective, based on the dynamic response characteristics of the aforementioned mechanism model, the desired reference trajectory is generated, and the reference trajectory is discretely sampled in terms of the number of time nodes to obtain the node-constrained target particle size.

[0050] In this embodiment, when analyzing the response characteristics of crushing and classification kinetics based on the mechanistic model, a stable operating point is selected, and small step disturbances are applied to the classifier speed, feed rate, or crushing air pressure one by one. The response curves of representative particle size statistics over time (such as D50 or the cumulative content of a certain threshold particle size) are recorded. The main response time scale is extracted using the half-amplitude time method, thereby obtaining the dominant time constant of the process. For example, after the classifier speed is stepped up, the half-amplitude time of D50 is measured to be 2s, then the dominant time constant is taken as 2s.

[0051] Then, based on the obtained dominant time constant, the prediction step size and prediction time domain for inverse optimization are determined, and the number of intermediate nodes is calculated. In this process, the prediction step size is set to an equal or proportional value to the dominant time constant. Combined with the given processing time, the covered time interval is determined. Then, the time axis is discretized at equal intervals according to the prediction step size. The total number of discretized time nodes is counted, and the number of intermediate nodes is obtained by subtracting the start and end points. For example, when the dominant time constant is 2s and the processing time is 20s, the prediction step size is 2s, the time nodes are 0~20s (11 in total), and the number of intermediate nodes is 9.

[0052] Finally, with the crushing compensation target particle size as the objective, a desired reference trajectory is generated based on the dynamic response characteristics of the mechanism model. Discrete sampling is then performed according to the aforementioned number of time nodes to obtain the node-constrained target particle size. Specifically, the reference trajectory employs exponential approach or spline interpolation to monotonically converge the particle size statistics from the initial value to the crushing compensation target particle size. At each discrete time node, the target particle size value corresponding to the reference trajectory is read, forming a one-to-one correspondence set between time nodes and target particle sizes. For example, with an initial D50 of 12 μm, a target of 8 μm, and a prediction step size consistent with the example, a sequence of target values ​​gradually approximating 8 μm is obtained at nodes 2, 4, 6…20 s, thereby obtaining the node-constrained target particle size.

[0053] Furthermore, in the method provided in the application embodiment, the forward sequence is flipped according to the inverse control parameter sequence, and parameter verification is performed with the collected particle size distribution data as the initial step and the target particle size distribution as the final target. The method further includes: The crushing control parameter sequence obtained by reverse deduction is arranged in forward time sequence; with the collected particle size distribution data as the initial state, the forward-arranged control parameter sequence is input into the updated hybrid digital twin model for forward simulation to obtain the matching degree between the final particle size distribution of the verification simulation results and the target particle size distribution; when the matching degree reaches the preset threshold, the control parameter sequence is confirmed to be effective and the parameter verification is passed.

[0054] In this embodiment, the reverse control parameter sequence is first reversed in chronological order, and then the parameter sequence from the end to the beginning is converted into a time sequence from the beginning to the end based on the control action sequence obtained through reverse deduction. For example, if the original sequence is t3→t2→t1→t0, the corresponding forward sequence is t0→t1→t2→t3. This process yields a forward-ordered control parameter sequence.

[0055] The collected particle size distribution data is then used as the initial state, and the forward-aligned sequence of control parameters is input into the updated hybrid digital twin model for forward simulation. At each time step, the corresponding stager speed, feed rate, and crushing air pressure are applied to the model. The crushing rate function and staged efficiency function from the mechanistic side are used to complete the numerical progression, obtaining the simulated final particle size distribution. For example, with a step size of 2s and a total time domain of 20s, parameters are loaded sequentially at t=0, 2, ..., 20s, and the final particle size distribution is output for verification.

[0056] Finally, the consistency between the simulated final particle size distribution and the target particle size distribution is evaluated, and the matching degree is calculated to determine whether the parameter verification has passed. In this process, the normalized mean square error method is used, where the squared differences between each particle size group are summed and normalized to obtain the error E. The matching degree = 1 − E represents the consistency and is compared with a preset threshold. For example, if E = 0.03, the matching degree is 0.97; if the threshold is 0.95, the control parameter sequence is considered valid. When the matching degree reaches the preset threshold, the control parameter sequence is confirmed to be valid, and parameter verification is completed.

[0057] In summary, the embodiments of this application have at least the following technical effects: This application establishes a hybrid digital twin model, including a mechanistic model and a data-driven model; connects to a particle size monitoring network, establishes particle size distribution data according to the collected monitoring distribution relationship, and obtains process parameters, including at least the classifier wheel speed, feed rate, and pulverizing air pressure; inputs the process parameters into the mechanistic model to simulate the pulverizing state, and iteratively updates the model parameters through the data-driven model based on the particle size distribution data, interactively updating the parameters of the hybrid digital twin model; performs differential analysis based on the real-time monitored particle size distribution data and the target particle size distribution to determine the pulverizing compensation target particle size; establishes a reverse optimization model based on the pulverizing compensation target particle size and the updated hybrid digital twin model, performs adaptive search of reverse pulverizing parameters according to the pulverizing compensation target particle size, obtains pulverizing control parameters, and feeds them back to the controller for adaptive control of the pulverizing airflow. This invention addresses the technical problems of large particle size deviation and poor stability in the process of air jet milling of pharmaceuticals in existing technologies. By establishing a hybrid digital twin model that combines a mechanistic model and a data-driven model, and performing differential analysis and reverse optimization parameter search based on real-time particle size distribution data, this invention achieves real-time, effective, and precise control of pharmaceutical air jet milling using an intelligent control system. It also considers the specific environmental characteristics of particle size distribution, thereby improving the adaptability of milling control and the robustness of the system.

[0058] Example 2 is based on the same inventive concept as the adaptive modeling and control method for drug airflow pulverization based on particle size distribution in the previous examples, such as... Figure 2 As shown, this application provides an adaptive modeling and control system for drug airflow pulverization based on particle size distribution. The system and method embodiments in this application are based on the same inventive concept. The system includes: The model building module 11 is used to build a hybrid digital twin model, including a mechanistic model and a data-driven model; the monitoring module 12 is used to connect to the particle size monitoring network, build particle size distribution data according to the collected monitoring distribution relationship, and obtain process parameters, including at least the classifier wheel speed, feed rate, and pulverizing air pressure; the parameter update module 13 is used to input the process parameters into the mechanistic model to simulate the pulverizing state, and to iteratively update the model parameters through the data-driven model based on the particle size distribution data, interactively updating the parameters of the hybrid digital twin model; the differential analysis module 14 is used to perform differential analysis based on the real-time monitored particle size distribution data and the target particle size distribution to determine the pulverizing compensation target particle size; the adaptive control module 15 is used to build a reverse optimization model based on the pulverizing compensation target particle size and the updated hybrid digital twin model, to perform an adaptive search of reverse pulverizing parameters according to the pulverizing compensation target particle size, to obtain pulverizing control parameters, and to feed them back to the controller for adaptive control of the pulverizing airflow.

[0059] Furthermore, the system is also used to implement the following functions: Based on the operating mechanism of the pulverizer and sample experimental data, a mechanistic model is constructed to characterize the crushing and classification dynamics of particles during airflow pulverization. A data-driven model is trained based on a neural network framework to perform real-time correction of the mechanistic model. Its input data is real-time process data, and its output is the compensation values ​​for the parameters in the mechanistic model. The physical structure of the pulverizer is collected and combined with the mechanistic model to establish a digital twin model. The data-driven model is used as an internal learning module to establish an interactive learning path with the mechanistic model, and the hybrid digital twin model is obtained by integration. The integration method is to dynamically inject the output of the data-driven module as real-time updated parameters into the mechanistic model.

[0060] Furthermore, the system is also used to implement the following functions: The mechanism model includes a crushing rate function and a grading efficiency function. The data-driven model is a data-driven correction model. The model output compensation value is the dynamic compensation amount of the key function parameters in the mechanism model, including the crushing rate compensation amount and the grading efficiency compensation amount.

[0061] Furthermore, the system is also used to implement the following functions: The real-time collected process parameters are used as input to the mechanism model to simulate the state of the crushing process, and the predicted particle size distribution under the current working conditions is output. The predicted particle size distribution is compared with the real-time collected particle size distribution data to obtain the particle size distribution difference. Based on the particle size distribution difference, the model parameters of the data-driven model are updated online using a recursive estimation algorithm, so that the output compensation value of the data-driven model can reduce the particle size distribution difference. The search is iterated until the error threshold is met, and the optimal model parameters are determined to update the parameters of the mechanism model in the hybrid digital twin model.

[0062] Furthermore, the system is also used to implement the following functions: Based on the updated hybrid digital twin model, a reverse optimization model is established with the crushing compensation target particle size as the control target. Based on the mechanism model, reverse simulation is performed to obtain the reverse control parameter sequence. According to the reverse control parameter sequence, the forward sequence is reversed, and the parameters are verified with the collected particle size distribution data as the initial step and the target particle size distribution as the final target. When the parameter verification requirements are met, the crushing control parameters are determined.

[0063] Furthermore, the system is also used to implement the following functions: A control time sequence chain is constructed with the crushing process and processing time as constraints and the crushing compensation target particle size as the objective. Based on the mechanistic model, the control time sequence chain is divided into stages with constraints, the node constraint target particle size is determined, and the nodes are labeled. Based on the control time sequence chain and the node constraint target particle size, a probabilistic search algorithm is used for reverse deduction. Candidate parameter sequences are evaluated through multiple forward simulation iterations, and the optimal control parameter sequence is selected based on the deduction results.

[0064] Furthermore, the system is also used to implement the following functions: Using the current state as the root node and the final state of the time-series chain as the target, a probabilistic graph search tree is initialized. Nodes in the search tree represent states at specific times, and edges represent candidate control actions connecting states. Starting from the root node, the optimal child node is recursively selected based on the upper confidence interval algorithm until a scalable non-final state node is reached. One or more new child nodes are added to the selected node, each corresponding to a candidate control action that has not been fully explored. Starting from the newly expanded node, simulation is performed based on the hybrid digital twin model to the final state of the time-series chain, resulting in a complete state prediction trajectory. The reward value of the prediction trajectory is calculated, and the reward value comprehensively evaluates the degree of achievement of the final state target and the degree of constraint violation of intermediate nodes. The statistical information of all visited nodes is updated in reverse along the search path. After a predetermined number of iterations, the path with the highest comprehensive reward value is selected from the search tree, and the control action sequence corresponding to the path is extracted to obtain the inverse control parameter sequence.

[0065] Furthermore, the system is also used to implement the following functions: Based on the aforementioned mechanism model, the response characteristics of crushing dynamics and classification dynamics are analyzed, and the dominant time constant of the process is extracted. Based on the dominant time constant, the prediction step size and prediction time domain of the inverse optimization model are determined, and the number of intermediate nodes is calculated. Taking the crushing compensation target particle size as the objective, based on the dynamic response characteristics of the aforementioned mechanism model, the desired reference trajectory is generated, and the reference trajectory is discretely sampled in terms of the number of time nodes to obtain the node-constrained target particle size.

[0066] Furthermore, the system is also used to implement the following functions: The crushing control parameter sequence obtained by reverse deduction is arranged in forward time sequence; with the collected particle size distribution data as the initial state, the forward-arranged control parameter sequence is input into the updated hybrid digital twin model for forward simulation to obtain the matching degree between the final particle size distribution of the verification simulation results and the target particle size distribution; when the matching degree reaches the preset threshold, the control parameter sequence is confirmed to be effective and the parameter verification is passed.

[0067] It should be noted that the order of the embodiments described above is for descriptive purposes only and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0068] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0069] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. An adaptive modeling and control method for pharmaceutical airflow pulverization based on particle size distribution, characterized in that, include: Establish a hybrid digital twin model, including a mechanistic model and a data-driven model; Connect the particle size monitoring network, establish particle size distribution data according to the collected monitoring distribution relationship, and obtain process parameters, including at least the classifier wheel speed, feed rate, and pulverizing air pressure; The process parameters are input into the mechanism model to simulate the crushing state. Based on the particle size distribution data, the model parameters are iteratively updated through data-driven modeling. The parameters of the hybrid digital twin model are updated interactively. Based on the real-time monitored particle size distribution data, differential analysis is performed in conjunction with the target particle size distribution to determine the target particle size for crushing compensation. Based on the target particle size for crushing compensation and the updated hybrid digital twin model, an inverse optimization model is established. An adaptive search for inverse crushing parameters is performed according to the target particle size for crushing compensation to obtain crushing control parameters, which are then fed back to the controller for adaptive control of the crushing airflow.

2. The adaptive modeling and control method for drug airflow pulverization based on particle size distribution according to claim 1, characterized in that, Establish a hybrid digital twin model, including a mechanistic model and a data-driven model, including: Based on the operating mechanism of the pulverizer and combined with sample experimental data, a mechanism model was constructed to characterize the crushing dynamics and classification dynamics of particles in the air jet pulverization process. A data-driven model based on a neural network framework is used to perform real-time correction of the mechanism model. Its input data is real-time process data, and its output is the compensation value of the parameters in the mechanism model. A digital twin model is established by combining the physical structure of the crusher with the aforementioned mechanism model. The data-driven model is used as an internal learning module to establish an interactive learning path with the mechanism model. The hybrid digital twin model is obtained by integrating the data-driven module's output as a real-time updated parameter and dynamically injecting it into the mechanism model.

3. The adaptive modeling and control method for drug airflow pulverization based on particle size distribution according to claim 2, characterized in that, The mechanism model includes a crushing rate function and a grading efficiency function. The data-driven model is a data-driven correction model. The model output compensation value is the dynamic compensation amount of the key function parameters in the mechanism model, including the crushing rate compensation amount and the grading efficiency compensation amount.

4. The adaptive modeling and control method for drug airflow pulverization based on particle size distribution according to claim 3, characterized in that, The process parameters are input into the mechanism model to simulate the crushing state. Based on the particle size distribution data, the model parameters are iteratively updated using a data-driven model. The parameters of the hybrid digital twin model are updated interactively, including: The real-time collected process parameters are used as input to the mechanism model to simulate the state of the crushing process and output the predicted particle size distribution under the current working conditions. The predicted particle size distribution is compared with the real-time collected particle size distribution data to obtain the particle size distribution difference. Based on the particle size distribution difference, the model parameters of the data-driven model are updated online using a recursive estimation algorithm, so that the output compensation value of the data-driven model can reduce the particle size distribution difference. The search is iterated until the error threshold is met, and the optimal model parameters are determined to update the parameters of the mechanism model in the hybrid digital twin model.

5. The adaptive modeling and control method for drug airflow pulverization based on particle size distribution according to claim 4, characterized in that, Based on the target particle size for crushing compensation and the updated hybrid digital twin model, an inverse optimization model is established. An adaptive search for inverse crushing parameters is performed according to the target particle size for crushing compensation to obtain crushing control parameters, including: Based on the updated hybrid digital twin model, with the crushing compensation target particle size as the control target, an inverse optimization model is established. Based on the mechanism model, inverse simulation deduction is performed to obtain the inverse control parameter sequence. Based on the reverse control parameter sequence, the forward sequence is flipped, and parameter verification is performed with the collected particle size distribution data as the initial step and the target particle size distribution as the final target. When the parameter verification requirements are met, the crushing control parameters are determined.

6. The adaptive modeling and control method for drug airflow pulverization based on particle size distribution according to claim 5, characterized in that, Based on the updated hybrid digital twin model, and using the crushing compensation target particle size as the control objective, an inverse optimization model is established. Inverse simulation is then performed based on the aforementioned mechanism model to obtain the inverse control parameter sequence, including: A control timing chain is constructed with the crushing process and processing time as constraints and the crushing compensation target particle size as the objective. Based on the aforementioned mechanism model, the control time-series chain is segmented into stages to determine the target particle size of node constraints, and node annotation is performed. Based on the control timing chain and the node constraint target particle size, a probabilistic search algorithm is used for reverse deduction. The candidate parameter sequence is evaluated through multiple forward simulation iterations, and the optimal control parameter sequence is selected based on the deduction results.

7. The adaptive modeling and control method for drug airflow pulverization based on particle size distribution according to claim 6, characterized in that, The reverse deduction is performed using a probabilistic search algorithm, including: With the current state as the root node and the final state of the time-series chain as the target, initialize a probabilistic graph search tree. The nodes in the search tree represent the states at a specific time, and the edges represent candidate control actions connecting the states. Starting from the root node, the optimal child node is recursively selected based on the upper confidence interval algorithm until a scalable non-final state node is reached. Add one or more new child nodes to the selected node, each child node corresponding to a candidate control action that has not been fully explored; Starting from the newly expanded node, a complete state prediction trajectory is obtained by simulating to the final state of the time series based on the hybrid digital twin model. The reward value of the predicted trajectory is calculated. The reward value is used to comprehensively evaluate the degree of achievement of the final state goal and the degree of violation of the constraints of intermediate nodes. The statistical information of all visited nodes is updated in reverse along the search path. After a predetermined number of iterations, the path with the highest comprehensive reward value is selected from the search tree, and the control action sequence corresponding to the path is extracted to obtain the inverse control parameter sequence.

8. The adaptive modeling and control method for drug airflow pulverization based on particle size distribution according to claim 6, characterized in that, Based on the aforementioned mechanism model, the control time-series chain is segmented into stages to determine the target granularity of node constraints, including: Based on the aforementioned mechanism model, the response characteristics of crushing dynamics and staged dynamics are analyzed, and the dominant time constant of the process is extracted. Based on the dominant time constant, the prediction step size and prediction time domain of the inverse optimization model are determined, and the number of intermediate nodes is calculated. With the target particle size for crushing compensation as the objective, a desired reference trajectory is generated based on the dynamic response characteristics of the aforementioned mechanism model. The reference trajectory is then discretely sampled at different time points to obtain the node-constrained target particle size.

9. The adaptive modeling and control method for drug airflow pulverization based on particle size distribution according to claim 5, characterized in that, Based on the inverse control parameter sequence, the forward sequence is reversed, and parameter verification is performed using the collected particle size distribution data as the initial step and the target particle size distribution as the final target. This includes: The crushing control parameter sequence obtained by reverse deduction is arranged in forward time sequence; Using the collected particle size distribution data as the initial state, the forward-aligned sequence of control parameters is input into the updated hybrid digital twin model for forward simulation to obtain the degree of matching between the final particle size distribution and the target particle size distribution in verifying the simulation results. When the matching degree reaches the preset threshold, the control parameter sequence is confirmed to be valid and the parameter verification is passed.

10. An adaptive modeling and control system for pharmaceutical airflow pulverization based on particle size distribution, characterized in that, The system is used to execute the adaptive modeling and control method for drug airflow pulverization based on particle size distribution as described in any one of claims 1-9, and the system includes: The model building module is used to build hybrid digital twin models, including mechanistic models and data-driven models; The monitoring module is used to connect to the particle size monitoring network, establish particle size distribution data according to the collected monitoring distribution relationship, and acquire process parameters, including at least the classifier wheel speed, feed rate, and crushing air pressure. The parameter update module is used to input the process parameters into the mechanism model to simulate the crushing state, and to perform iterative updates of the model parameters through data-driven model based on the particle size distribution data, and to interactively update the parameters of the hybrid digital twin model. The differential analysis module is used to perform differential analysis based on the real-time monitored particle size distribution data and the target particle size distribution to determine the target particle size for crushing compensation. The adaptive control module is used to establish a reverse optimization model based on the target particle size for crushing compensation and the updated hybrid digital twin model, perform an adaptive search for reverse crushing parameters according to the target particle size for crushing compensation, obtain crushing control parameters, and feed them back to the controller for adaptive control of the crushing airflow.

Images