A method and device for optimizing process parameters of a traditional Chinese medicine tablet

Abstract

The present disclosure relates to a traditional Chinese medicine tablet manufacturing process parameter optimization method and device. The optimization method comprises: constructing a prediction model at each process stage of traditional Chinese medicine tablet manufacturing to establish the mapping relationship between process parameters and quality indicators, wherein the prediction model adaptively weights process parameter features through a channel attention mechanism and maintains gradient stability using a residual connection structure; using the output of the prediction model as the fitness evaluation benchmark, and using an improved whale optimization algorithm to search for the optimal parameter combination in the process parameter space, wherein the improved whale optimization algorithm includes dynamically shrinking the search range to realize the transition from global exploration to local fine search, applying normal distribution disturbance to enhance the local optimal escape ability, and retaining the historical optimal solution to guide the search direction; based on the optimal parameter combination, traditional Chinese medicine tablets are produced, and it is detected whether the actual quality indicators meet the preset target.

Description

TECHNICAL FIELD

[0001] The present disclosure belongs to the technical field of traditional Chinese medicine tablet manufacturing, and particularly relates to a method and device for optimizing process parameters of traditional Chinese medicine tablet manufacturing. BACKGROUND

[0002] The manufacturing process of traditional Chinese medicine tablets includes multiple continuous links such as granulation, mixing, and tablet pressing. There is a significant nonlinear coupling relationship between the process parameters of each stage, and the parameter settings of multiple process stages jointly affect the key quality attributes of the finished product. Quality indicators such as hardness, friability, tensile strength, tablet weight consistency, and disintegration time are not only affected by single-stage factors, but also limited by the complex interaction between cross-stage parameters.

[0003] Traditionally, the process parameters of each process stage in the production process are set based on experience, which requires a large number of experiments to determine the optimal parameter combination, which is time-consuming and labor-intensive. Currently, methods for quickly determining the optimal process parameter combination through neural network models have been applied in the field of traditional Chinese medicine tablet manufacturing. The neural network model learns the data-driven relationship between process parameters and quality responses at each stage to predict the optimal process parameter combination.

[0004] However, there are numerous process parameters at each stage of the traditional Chinese medicine tablet manufacturing process, making data collection difficult and costly, resulting in insufficient training data. The neural network model based on small sample training is prone to overfitting or unstable prediction, which affects the prediction accuracy; at the same time, although the neural network model can predict the quality response, it lacks an efficient reverse optimization mechanism, making it difficult to actively find the optimal parameter combination, and process control still requires human intervention, which is inefficient and costly; in addition, most neural network models are "black box structures" and lack the ability to explain the causal relationship between key process parameters and quality attributes, making it difficult to support scientific decision-making and quality traceability, which does not meet the requirements of controllable quality and traceable source in traditional Chinese medicine manufacturing. SUMMARY

[0005] The embodiments of the present disclosure propose an optimization scheme for process parameters of traditional Chinese medicine tablet manufacturing to solve the problems of insufficient prediction accuracy, the need for human intervention in process control, and the lack of model interpretability of existing optimization schemes based on neural network models under small sample training conditions.

[0006] A first aspect of the embodiments of the present disclosure provides a method for optimizing process parameters of traditional Chinese medicine tablet manufacturing, comprising:

[0007] A prediction model is constructed at each process stage of traditional Chinese medicine tablet manufacturing to establish a mapping relationship between process parameters and quality indicators, wherein the prediction model adaptively weights process parameter features through a channel attention mechanism and maintains gradient stability using a residual connection structure;

[0008] The output of the prediction model is taken as a fitness evaluation benchmark, and an improved whale optimization algorithm is used to search for an optimal parameter combination in a process parameter space, wherein the improved whale optimization algorithm includes dynamic contraction of a search range to realize a transition from global exploration to local fine search, application of normal distribution disturbance to enhance local optimal escape ability, and reservation of a historical optimal solution to guide a search direction;

[0009] Based on the optimal parameter combination, traditional Chinese medicine tablet production is performed, and it is detected whether an actual quality index reaches a preset target.

[0010] In some embodiments of the present disclosure, a SHAP method is introduced to perform explanatory analysis on a prediction result of the prediction model, so as to quantitatively reveal marginal contributions of various process variables in a traditional Chinese medicine manufacturing process to tablet quality.

[0011] In some embodiments of the present disclosure, an implementation method of the channel attention mechanism includes:

[0012] Global average pooling and maximum pooling are performed in parallel on input features to obtain first statistical features and second statistical features, respectively;

[0013] The first statistical features and the second statistical features are input into a weight generation network composed of two convolution layers, wherein a first convolution layer compresses an original channel number to K, and a second convolution layer restores to the original channel number C, wherein, K is a compressed channel number, C is an original channel number, is a compression ratio factor;

[0014] Features output by the first convolution layer and the second convolution layer are element-wise added, and channel attention weights are generated through a Sigmoid activation function;

[0015] The channel attention weights are multiplied by the input features to realize adaptive enhancement of a key process parameter channel.

[0016] In some embodiments of the present disclosure, the compression ratio factor is 4, 8, or 16;

[0017] The number of neurons in the last layer of the weight generation network is 10, 20, 40, 60, 80, or 100.

[0018] In some embodiments of the present disclosure, the residual connection structure includes:

[0019] A main path processing unit composed of at least one fully connected layer is used to perform nonlinear transformation on input features;

[0020] A jump connection unit composed of a fully connected layer is used to linearly project input features to a main path output dimension;

[0021] an output unit configured to add the output of the skip connection unit and the output of the main path processing unit element by element to generate a final output of the residual connection structure.

[0022] In some embodiments of the present disclosure, the number of neurons of the last fully connected layer in the main path processing unit is 10, 20, 40, 60, 80, or 100;

[0023] The number of neurons of the fully connected layer of the skip connection unit is the same as the first layer of the main path processing unit;

[0024] The number of residual connection structures is 2.

[0025] In some embodiments of the present disclosure, the dynamic contraction of the search range to achieve the transition from global exploration to local fine search includes:

[0026] The control parameter is dynamically adjusted based on the following formula to achieve adaptive adjustment of search capability:

[0027] ,

[0028] wherein, is a control parameter for controlling the search radius of the improved whale optimization algorithm in a multi-dimensional parameter space, is an initial control parameter, denotes the current number of iterations, is the maximum number of iterations, is a decay coefficient, and e is the base of the natural logarithm.

[0029] In some embodiments of the present disclosure, the application of normal distribution disturbance to enhance the local optimal escape ability includes:

[0030] After the individual position is updated, a random disturbance is applied to the current optimal solution, and the disturbed position is:

[0031] ,

[0032] wherein, is the position of the current optimal solution, is the position of the current optimal solution after disturbance, is the disturbance amplitude, denotes a standard normal distribution.

[0033] In some embodiments of the present disclosure, the reservation of historical optimal solutions to guide the search direction includes:

[0034] In the process of each generation iteration, the current historical optimal individual is reserved as an elite individual, and the elite individual is directly introduced into the next generation population for updating.

[0035] A second aspect of the embodiments of the present disclosure provides a Chinese medicine tablet manufacturing process parameter optimization device, comprising:

[0036] A modeling module is configured to construct a prediction model at each process stage of Chinese medicine tablet manufacturing to establish a mapping relationship between process parameters and quality indicators, wherein the prediction model is configured to adaptively weight process parameter features through a channel attention mechanism and maintain gradient stability by using a residual connection structure.

[0037] An optimization module is configured to use the output of the prediction model as a fitness evaluation benchmark and search for an optimal parameter combination in a process parameter space by using an improved whale optimization algorithm, wherein the improved whale optimization algorithm includes dynamically shrinking a search range to realize the transition from global exploration to local fine search, applying normal distribution disturbance to enhance the local optimal escape ability, and retaining a historical optimal solution to guide the search direction.

[0038] A verification module is configured to produce Chinese medicine tablets based on the optimal parameter combination and detect whether the actual quality indicators meet the preset target.

[0039] In summary, the Chinese medicine tablet manufacturing process parameter optimization method and device provided by the embodiments of the present disclosure improve the training stability of network depth through a residual structure, automatically identify key process variables (such as main compression strength and lubricant dosage) by introducing a channel attention mechanism, and ultimately achieve high prediction accuracy under the condition of small sample training. At the same time, by integrating the improved intelligent optimization algorithm IMWOA, the device realizes intelligent search of the optimal process parameter combination driven by the quality target in the multi-objective space, and adaptively escapes from the local optimal ability and global search stability, thereby avoiding manual intervention, significantly improving the parameter design efficiency and reducing the debugging cost. At the same time, by introducing the SHAP explainability analysis method, the marginal contribution of each process parameter to the target quality attribute (such as hardness, tensile strength, and disintegration time limit) is quantitatively evaluated, the model transparency and process mechanism understanding ability are enhanced, which is helpful to support the quality traceability and risk control in the Chinese medicine manufacturing process. BRIEF DESCRIPTION OF DRAWINGS

[0040] The features and advantages of the present disclosure will be more clearly understood through reference to the accompanying drawings, which are schematic and should not be interpreted as limiting the present disclosure in any way, in which:

[0041] Figure 1 is a flowchart of a Chinese medicine tablet manufacturing process parameter optimization method according to some embodiments of the present disclosure;

[0042] Figure 2 is a structural schematic diagram of a CA-Res-MLP model in the present disclosure;

[0043] Figure 3A schematic diagram of an apparatus for optimizing manufacturing process parameters of a traditional Chinese medicine tablet according to some embodiments of the present disclosure. DETAILED DESCRIPTION

[0044] In the following detailed description, numerous specific details are set forth in order to provide a thorough understanding of the disclosure. However, it will be apparent to one of ordinary skill in the art that the disclosure can be practiced without these details. In other instances, well-known methods, procedures, components, and network architectures have not been described in detail so as not to unnecessarily obscure aspects of the disclosure.

[0045] It will be understood that when an apparatus, unit, or module is referred to as being "on", "connected to", or "coupled to" another apparatus, unit, or module, it can be directly on, connected, or coupled to the other apparatus, unit, or module, or intervening apparatuses, units, or modules can be present, unless the context clearly indicates otherwise. For example, as used in the disclosure, the term "and / or" includes any and all combinations of one or more of the associated listed items.

[0046] The terminology used in the disclosure herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the disclosure. As used in the specification and in the claims, the phrase "a", "an" and / or "the" can connote "one or more" unless the context clearly dictates otherwise. In general, the term "includes" and / or "containing" is used to indicate the inclusion of one or more features, elements, steps, operations, integers, and / or components, but does not exclude the presence of one or more other features, elements, steps, operations, integers, and / or components.

[0047] These and other features and characteristics of the disclosure, as well as the methods of operation and the functions of the related elements of the structure and the combination of parts and economies of manufacture, will become more apparent upon consideration of the following description and the appended claims with reference to the accompanying drawings, all of which form a part of this specification. It is to be expressly understood, however, that the drawings are for the purpose of illustration and description only and are not intended as a definition of the limits of the disclosure. As such, it should be readily appreciated that the drawings are not necessarily drawn to scale of the embodiments.

[0048] Various structural diagrams are used in the disclosure to illustrate various modifications according to embodiments of the disclosure. It should be understood that the foregoing or the following structures are not intended to limit the disclosure. The scope of protection of the disclosure is subject to the claims.

[0049] The typical process of Chinese medicine tablet manufacturing includes granulation, mixing, tabletting and other continuous links. There is a significant nonlinear coupling relationship between the process parameters of each stage, especially influenced by the fluctuation of the physical and chemical properties of complex components of traditional Chinese medicine raw materials (such as polysaccharides, flavonoids, saponins, volatile oils, etc.), resulting in a high degree of uncertainty in the entire manufacturing system. The parameter settings of multiple process stages will jointly affect the key quality attributes of the finished product, such as hardness, friability, tensile strength, tablet weight consistency, and disintegration time, etc. These quality indicators are not only influenced by single-stage factors, but also limited by the complex interaction between cross-stage parameters.

[0050] Traditionally, the process parameters of each process stage are set based on experience, requiring a large number of experiments to determine the optimal parameter combination, which is time-consuming and labor-intensive. Currently, methods using neural network models to quickly determine the optimal process parameter combination have been applied in the field of Chinese medicine tablet manufacturing. Based on neural network models, the relationship between process parameters and quality responses of each stage is learned through data-driven learning, which can quickly predict the optimal process parameter combination. However, the application of the method in the multi-stage manufacturing of Chinese medicine tablets still has the following key technical bottlenecks:

[0051] Dependence on large sample training, weak generalization ability: Neural network models have a large number of parameters, and in the context of Chinese medicine manufacturing, which is difficult to obtain samples and has high experimental costs, often face the problem of insufficient training, and are prone to overfitting or unstable prediction, limiting their engineering practicality.

[0052] Lack of intelligent optimization ability, parameter adjustment still relies on manual experience: Although the model can predict the quality response, it lacks an efficient reverse optimization mechanism, making it difficult to actively find the optimal parameter combination. Process control still requires human intervention, which is inefficient and costly.

[0053] Lack of interpretability, weak decision support: Most neural network models are "black box structures" and lack the ability to explain the causal relationship between key process parameters and quality attributes, making it difficult to support scientific decision-making and quality traceability, which does not meet the requirements of controllable quality and traceable source in Chinese medicine manufacturing.

[0054] In addition, the complexity of components, large batch differences, and detection lag of Chinese medicine tablet systems make the above problems more prominent in actual production.

[0055] In summary, the current modeling methods based on neural networks have significant shortcomings in small sample size, high nonlinear coupling, full-process collaborative optimization, and model interpretability.

[0056] To solve the above problems, the disclosure proposes an improved neural network model (CA-Res-MLP model) that fuses channel attention mechanism and residual structure, and cooperates with an improved whale migration optimization algorithm (IMWOA algorithm) to form a "two-way intelligent optimization framework", realizing the collaborative fusion of high-precision process-quality modeling and reverse parameter solving, and effectively breaking through the key bottlenecks of existing technologies in small samples, system coupling and optimization intelligence.

[0057] The traditional Chinese medicine tablet manufacturing unit is configured with a process parameter optimization server for determining the optimal combination of process parameters for production based on given traditional Chinese medicine tablet quality indicators. The optimization method of the traditional Chinese medicine tablet manufacturing process parameters is executed by the process parameter optimization server, as shown in Figure 1 The specific steps include the following steps:

[0058] S110, a prediction model is constructed at each process stage of traditional Chinese medicine tablet manufacturing to establish the mapping relationship between process parameters and quality indicators, wherein the prediction model adaptively weights the process parameter features through channel attention mechanism, and maintains gradient stability by using residual connection structure.

[0059] The disclosure first acquires the numerical values of the preset process parameters at each process stage of traditional Chinese medicine tablet manufacturing and the numerical values of the preset quality evaluation indicators obtained based on the process parameters, takes the numerical values of the process parameters as sample data, and takes the numerical values of the quality evaluation indicators as labels, and performs data-driven learning to construct a prediction model for establishing the mapping relationship between process parameters and quality indicators.

[0060] Specifically:

[0061] In the granulation link, the granulation temperature, sample injection speed, spray pressure, and binder concentration data are obtained as sample data, and the moisture content index is taken as the label of the sample data to construct training data. Table 1 is an example of training data in the granulation link.

[0062] (Table 1)

[0063]

[0064] In the mixing link, the mixing time, mixing speed, and lubricant dosage data are obtained as sample data, and the angle of repose index is taken as the label of the sample data to construct training data. Table 2 is an example of training data in the mixing link.

[0065] (Table 2)

[0066]

[0067] In the tabletting link, the main pressure, filling amount, tabletting speed, filling speed, and pre-pressure data are obtained as sample data, and the porosity, hardness, and tensile strength indicators are obtained as labels of the sample data to construct training data. Table 3 is an example of tabletting link training data.

[0068] (Table 3)

[0069]

[0070] The present disclosure adopts a multi-round cross-validation mechanism to improve the reliability of the model and enhance the generalization ability under small sample data. Specifically, a 20-round 4-fold cross-validation strategy is adopted. The execution process is as follows: (1) data division: divide the entire sample into 4 equal parts, and each time leave one fold as the validation set and the other 3 folds as the training set; (2) repeat 20 times: to reduce the influence of random division on the evaluation results, the entire 4-fold cross-validation process is repeated for 20 rounds, and the division order is shuffled each time; (3) average evaluation indicators: the final model performance such as RMSE and R² is output by the average value of 20 rounds, which is more stable and reliable; (4) select the best model parameters: combined with grid search, the neuron number and attention compression rate and other hyperparameters are jointly tuned to ensure the optimal structure.

[0071] The prediction model (referred to as CA-Res-MLP model) used in the present disclosure is based on the traditional multi-layer perceptron (MLP) structure, which integrates the channel attention mechanism (Channel Attention, CA) and the residual connection structure (Residual Block), significantly improving the modeling ability of the nonlinear, multivariate process parameters on the quality response relationship in traditional Chinese medicine tablet manufacturing.

[0072] As shown in Figure 2 , the CA-Res-MLP model includes a channel attention mechanism and a residual connection structure. Specifically:

[0073] Channel attention mechanism:

[0074] The channel attention mechanism can perform weight redistribution on the input feature dimensions, effectively highlighting representative traditional Chinese medicine process variables (such as granulation temperature, binder concentration, lubricant dosage, and tabletting main pressure). The workflow is as follows:

[0075] (1) Perform global average pooling and maximum pooling on the input features respectively to extract two channel-level statistical features;

[0076] (2) Input the above two features into a network module composed of two convolutional layers to enhance the model's expression ability of channel importance; among them, the first convolutional layer performs dimensionality reduction compression on the original channel, and the second convolutional layer restores the channel number after dimensionality reduction compression to the original channel number;

[0077] (3) Element-level addition is performed on the extraction results of the two branches, and a channel attention weight is generated through a Sigmoid activation function;

[0078] (4) The weight is applied to the channel dimension of the original input feature to achieve adaptive enhancement of the key channel.

[0079] In the actual implementation process, the input feature is inserted with a dimension through the Unsqueeze 1D operation before entering the channel attention module, so that the information transformation acts on the channel dimension. This design enables the channel attention module to be seamlessly embedded in the residual structure, thereby effectively improving the feature extraction capability.

[0080] Two residual blocks that embed the channel attention mechanism are embedded in the hidden layer of the CA-Res-MLP model.

[0081] To further optimize the model performance, some embodiments of the disclosure use a grid search strategy to jointly tune two key hyperparameters: one is the number of neurons in the last layer of the residual main path (candidate values are 10, 20, 40, 60, 80, 100); the other is the compression ratio factor of the first convolutional layer in the channel attention mechanism (candidate values are 4, 8, 16).

[0082] The compression ratio factor is defined as follows:

[0083]

[0084] where K is the compressed channel number, C is the original channel number, is the compression ratio factor.

[0085] Residual connection structure:

[0086] To avoid the gradient vanishing caused by the increase of network depth, enhance the stability and feature expression ability of modeling, the CA-Res-MLP model adopts a residual connection structure (referred to as a residual block). Specifically, the residual block adopts a skip connection containing a fully connected layer, which directly transmits the input feature to the output end of the module, adds it to the nonlinear transformation output of the main path, and forms the final output. This design not only helps to maintain the transmission of original information, but also allows the network to autonomously choose whether to use deeper features during the training process according to the actual needs, thereby effectively alleviating the degradation problem in deep neural networks.

[0087] The two residual blocks are embedded in the hidden layers of the CA-Res-MLP model. The number of neurons in each module main path adopts a layer-by-layer decreasing setting strategy. In order to further optimize the performance of the model, the grid search method is used to adjust the number of hidden neurons in the last layer of the main path, and the candidate values are: 10, 20, 40, 60, 80 and 100. In an embodiment of the present disclosure, the settings of the remaining hyperparameters are as follows: the Dropout probability is 0.2, the number of training rounds is 120, the batch size is 4, the optimizer is Adam, and the learning rate is set to 0.001.

[0088] In S120, the output of the prediction model is used as the fitness evaluation benchmark, and an improved whale optimization algorithm is used to search for the optimal parameter combination in the process parameter space, wherein the improved whale optimization algorithm includes dynamic contraction of the search range to realize the transition from global exploration to local fine search, application of normal distribution disturbance to enhance the local optimal escape ability, and reservation of historical optimal solutions to guide the search direction.

[0089] In order to inversely deduce the process parameters for the quality target, the present disclosure proposes an improved whale migration optimization algorithm (IMWOA). In order to be applicable to the high-dimensional coupled parameter space in traditional Chinese medicine manufacturing, the IMWOA includes the following improvements compared to the standard whale migration optimization algorithm (WOA):

[0090] Dynamic control parameter mechanism:

[0091] The dynamic control parameter mechanism can adaptively adjust the search and development ability according to the search process, and improve the adaptability of the algorithm to the complex solution space of traditional Chinese medicine manufacturing. Specifically as follows: in the standard WOA, the parameter controls the search range, and its initial value is large and gradually decreases in the later period to realize the transition from global search to local development. The present disclosure uses an exponential decay method to dynamically adjust it, so that the algorithm has a wider search radius in the initial stage, which is beneficial to quickly jump out of the local optimal region, and in the later period, it gradually focuses on fine search near the high-quality solution. This mechanism enhances the stage search adaptability of the algorithm and effectively improves the overall optimization efficiency. The adjustment strategy is as follows:

[0092]

[0093] wherein, is a control parameter for controlling the search radius of the improved whale optimization algorithm in the multi-dimensional parameter space, is an initial control parameter, represents the current iteration number, is the maximum iteration number, is a decay coefficient, and e is the base of natural logarithm.

[0094] Adaptive disturbance mechanism:

[0095] To avoid the algorithm falling into local optimum in complex non-convex optimization space, the adaptive perturbation strategy is designed in the disclosure. Specifically, after the whale individual updates the position, a small perturbation is introduced to the current individual position according to a certain probability

[0096]

[0097] wherein, is the position of the current optimal solution, is the position of the current optimal solution after perturbation, is the perturbation amplitude, represents the standard normal distribution.

[0098] This strategy effectively breaks the population concentration trend under the premise of ensuring the perturbation control range, and enhances the exploration ability of the solution space and the possibility of jumping out of the local optimum.

[0099] Elite preservation mechanism:

[0100] The elite preservation mechanism can preserve the current population optimal solution and ensure the quality and search direction of the solution. Specifically, to improve the quality and stability of the solution, in each iteration, the algorithm preserves the current historical optimal individual and includes it in the next generation search process as a reference or seed individual to participate in the update. This strategy avoids the risk of losing excellent solutions due to random perturbation in the search process, and guides the population to converge to the high-quality solution direction, improving the convergence stability and prediction robustness of the algorithm.

[0101] S130, based on the optimal parameter combination, producing traditional Chinese medicine tablets, and detecting whether the actual quality index reaches the preset target.

[0102] Some embodiments of the disclosure also introduce the SHAP (SHapley Additive exPlanations) method to perform explanatory analysis on the model prediction results, to quantitatively reveal the marginal contribution of each process variable in the traditional Chinese medicine manufacturing process to the tablet quality, thereby providing a theoretical basis for drug quality traceability and process decision-making.

[0103] As shown in Figure 3 The traditional Chinese medicine tablet manufacturing process parameter optimization device 300 includes a modeling module 310, an optimization module 320, and a verification module 330. In some embodiments of the disclosure, the optimization function of the traditional Chinese medicine tablet manufacturing process parameters is executed by a process parameter optimization server, wherein:

[0104] The modeling module 310 is configured to construct a prediction model for each process stage of the traditional Chinese medicine tablet manufacturing to establish a mapping relationship between the process parameters and the quality indicators, wherein the prediction model is configured to perform self-adaptive weighting on the process parameter features through a channel attention mechanism, and adopt a residual connection structure to maintain gradient stability.

[0105] The optimization module 320 is configured to take the output of the prediction model as an adaptability evaluation benchmark, and search for an optimal parameter combination in a process parameter space by using an improved whale optimization algorithm, wherein the improved whale optimization algorithm includes dynamic shrinkage of a search range to realize a transition from global exploration to local fine search, application of a normal distribution disturbance to enhance local optimal escape ability, and reservation of a historical optimal solution to guide a search direction.

[0106] The verification module 330 is configured to produce the traditional Chinese medicine tablet based on the optimal parameter combination, and detect whether the actual quality indicators reach a preset target.

[0107] In summary, the method and device for optimizing the process parameters of the traditional Chinese medicine tablet manufacturing provided by the embodiments of the present disclosure improve the training stability of the network depth through the residual structure, automatically identify the key process variables (such as the main compression strength and the lubricant dosage) by introducing the channel attention mechanism, and finally achieve a high prediction accuracy under the condition of small sample training. Meanwhile, the intelligent optimization algorithm of the improved IMWOA is integrated to realize intelligent search for the optimal process parameter combination driven by the quality target in the multi-objective space, and the self-adaptive jumping out of the local optimal ability and the global search stability are realized, so as to avoid manual intervention, significantly improve the parameter design efficiency and reduce the debugging cost. Meanwhile, the SHAP explainability analysis method is introduced to quantitatively evaluate the marginal contribution of each process parameter to the target quality attribute (such as the hardness, tensile strength, and disintegration time limit), enhance the model transparency and process mechanism understanding ability, and help support the quality traceability and risk control in the traditional Chinese medicine manufacturing process.

[0108] Those skilled in the art can clearly understand the specific working process of the devices and modules described above for the convenience and brevity of description, which can be referred to the corresponding description in the foregoing device embodiments, and will not be described here.

[0109] Although the subject matter described herein is provided in the general context of computer-executable instructions of a program module being executed by a computer system on a computer-readable medium, those skilled in the art will recognize that those alternative implementations can be provided as well, including but not limited to, distributed system or component software programs, parallel processing systems, and other computational architectures where task and functionality can be implemented as one or more modules. Generally, program modules include routines, programs, components, data structures, and other types of structures that perform particular tasks or implement particular abstract data types. Those skilled in the art will also recognize that those subject matters described herein can be practiced with other computer system configurations, including hand-held devices, multiprocessor systems, microprocessor-based or programmable consumer electronics, minicomputers, mainframe computers, and the like. Those skilled in the art will recognize that the subject matter described herein can be practiced in distributed computing environments where tasks are performed by remote processing devices that are linked through a communications network. In this case, program modules can be located in both local and remote memory storage devices.

[0110] Those skilled in the art will realize that the exemplary units and method steps described in connection with the embodiments disclosed herein can be carried out by electronic hardware, computer software, or combinations of both. Whether such functionality is implemented in hardware or software depends on the particular applications and design constraints imposed on the overall system. Those skilled in the art can implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present disclosure.

[0111] It should be understood that the foregoing detailed description of the disclosure, rather than limiting the disclosure, is intended to provide an overview or framework for understanding the subject matter. Therefore, any modification, equivalent replacement, improvement, etc. made without departing from the spirit and scope of the disclosure should be included in the protection scope of the disclosure. In addition, the claims of the disclosure are intended to cover all variations and modifications falling within the scope and boundary of the appended claims or the equivalent forms of such scope and boundary.

Claims

1. A method for optimizing the process parameters for manufacturing Chinese medicine tablets, characterized in that, The method comprises the following steps: A prediction model is constructed in the process stage of traditional Chinese medicine tablet manufacturing to establish the mapping relationship between process parameters and quality indicators, and the process stage at least includes the stages of granulation, mixing and tabletting, the process data has small sample characteristics, and there is a nonlinear coupling relationship between the parameters, wherein, for the small sample characteristics and nonlinear coupling relationship between the process parameters, the prediction model adaptively weights the features of one-dimensional process parameter sequence through a channel attention mechanism, wherein the channel attention mechanism processes the one-dimensional input sequence through global average pooling, adapts the channel dimension through Unsqueeze 1D operation, and maintains gradient stability through a residual connection structure, the channel attention mechanism and the residual connection structure constitute a CA-Res-MLP model, and the prediction model is trained through a multi-round cross-validation mechanism to solve the parameter optimization and overfitting problems under the small sample training condition; With the output of the prediction model as a fitness evaluation benchmark, an improved whale optimization algorithm is used to search for an optimal parameter combination in a process parameter space, the improved whale optimization algorithm including dynamic contraction of a search range based on an exponential decay law to achieve a transition from global exploration to local fine search, application of normal distribution disturbance to enhance local optimal escape ability, and reservation of a historical optimal solution to guide a search direction, the dynamic contraction of the search range based on the exponential decay law referring to contraction of a search range based on an exponential decay law Adaptive adjustment of search capability is achieved, wherein, a control parameter for controlling a search radius of the improved whale optimization algorithm in a multi-dimensional parameter space, an initial control parameter, represents a current iteration number, a maximum iteration number, a decay coefficient, e is the base of natural logarithm; Based on the optimal parameter combination, traditional Chinese medicine tablets are produced, and it is detected whether the actual quality indicators reach the preset target.

2. The method of claim 1, wherein, Further comprising: The SHAP method is introduced to perform explanatory analysis on the prediction results of the prediction model to quantitatively reveal the marginal contribution of each process variable in the traditional Chinese medicine manufacturing process to the tablet quality.

3. The method of claim 1, wherein, The implementation method of the channel attention mechanism comprises: The global average pooling and the maximum pooling are performed in parallel on the input features to obtain first statistical features and second statistical features respectively; The first statistical feature and the second statistical feature are respectively input into a weight generation network composed of two convolution layers, wherein a first convolution layer compresses the original channel number to K, and a second convolution layer restores to the original channel number C, wherein, K is the compressed channel number, C is the original channel number, is a compression ratio factor; The features output by the first convolutional layer and the second convolutional layer are element-wise added, and the channel attention weight is generated through a Sigmoid activation function; The channel attention weight is multiplied by the input features to realize adaptive enhancement of the key process parameter channel.

4. The method of claim 3, wherein: The compression scale factor is 4, 8 or 16; The number of neurons in the last layer of the weight generation network is 10, 20, 40, 60, 80 or 100.

5. The method of claim 1, wherein, The residual connection structure comprises: A main path processing unit composed of at least one fully connected layer, used for nonlinear transformation of input features; A jump connection unit composed of a fully connected layer, used for linear projection of input features to the output dimension of the main path; An output unit for element-wise addition of the output of the jump connection unit and the output of the main path processing unit to generate the final output of the residual connection structure.

6. The method of claim 5, wherein: The number of neurons in the last layer of the main path processing unit is 10, 20, 40, 60, 80 or 100; The number of neurons of the fully connected layer of the jump connection unit is the same as that of the first layer of the main path processing unit; The number of residual connection structures is 2.

7. The method of claim 1, wherein, The step of applying normal distribution disturbance to enhance the local optimal escape ability comprises: After updating the individual position, a random disturbance is applied to the current optimal solution, and the disturbed position is: ， wherein, is the position of the current optimal solution, is the position of the current optimal solution after perturbation, is the perturbation amplitude, denotes the standard normal distribution.

8. The method of claim 1, wherein, The step of retaining the historical optimal solution to guide the search direction comprises: In each generation iteration process, the current historical optimal individual is retained as an elite individual, and the elite individual is directly introduced into the next generation population for updating.

9. A device for optimizing the process parameters of Chinese medicine tablet manufacturing, characterized in that, The method comprises the following steps: The modeling module is used to construct a prediction model at a process stage of traditional Chinese medicine tablet manufacturing to establish a mapping relationship between process parameters and quality indicators, the process stage at least including a granulation, mixing and tabletting stage, the process data having small sample characteristics, and there is a nonlinear coupling relationship between the parameters, wherein, for the small sample characteristics and nonlinear coupling relationship between the process parameters, the prediction model adaptively weights the features of one-dimensional process parameter sequence through a channel attention mechanism, wherein the channel attention mechanism processes one-dimensional input sequence through global average pooling, and adopts Unsqueeze 1D operation to adapt the channel dimension, and adopts a residual connection structure to maintain gradient stability, the channel attention mechanism and the residual connection structure constitute a CA-Res-MLP model, and the prediction model is trained through a multi-round cross-validation mechanism to solve the parameter optimization and overfitting problem under the small sample training condition; An optimization module is configured to search for an optimal parameter combination in a process parameter space using an improved whale optimization algorithm with the output of the prediction model as a fitness evaluation benchmark, the improved whale optimization algorithm including dynamic contraction of a search range based on an exponential decay law to achieve a transition from global exploration to local fine search, application of normal distribution disturbance to enhance local optimal escape ability, and reservation of a historical optimal solution to guide a search direction, the dynamic contraction of the search range based on the exponential decay law being based on Adaptive adjustment of search capability is achieved, wherein, is a control parameter for controlling the search radius of the improved whale optimization algorithm in a multi-dimensional parameter space, is an initial control parameter, represents a current iteration number, is a maximum iteration number, is a decay coefficient, and e is the base of natural logarithm. The verification module is used to produce traditional Chinese medicine tablets based on the optimal parameter combination, and detect whether the actual quality indicators reach the preset target.

Images