A method for predicting the content uniformity of cartilage powder tablets based on machine learning

Abstract

The application discloses a kind of cartilage powder tableting content uniformity prediction methods based on machine learning, it is related to pharmaceutical preparation optimization technical field, including, acquisition multidimensional cartilage powder tableting data, and the pre-processing of multidimensional cartilage powder tableting data is generated tabletting feature dataset;Through difference method, the time series error of uniformity prediction value is backtracked, obtains uniformity deviation, and adjusts the compaction force and temperature and humidity condition in uniformity deviation, generates uniformity error range;According to uniformity error range, the uniformity deviation of production parameter is adjusted by reward mechanism, and the compaction force and temperature and humidity process condition are optimized, and generate optimization control scheme.The application improves the precision and calculation efficiency of uniformity prediction, provides reliable machine learning method support for the optimization of cartilage powder tableting process.

Description

Technical Field

[0001] This invention relates to the field of pharmaceutical formulation optimization technology, and in particular to a method for predicting the uniformity of cartilage powder content in tablets based on machine learning. Background Technology

[0002] With the rapid development of the pharmaceutical industry, the production process of drug formulations increasingly emphasizes quality control and precision. Cartilage powder tableting, as a common drug formulation method, is widely used in therapeutic areas such as cartilage repair and regeneration due to its ability to precisely control drug content and quality. In traditional cartilage powder tableting production, uniformity control has always been a critical issue. Uniformity not only affects the efficacy of cartilage powder but also relates to the stability of the production process and the quality of the final product. How to accurately predict and control the uniformity of cartilage powder content in tablets has become a research hotspot in the pharmaceutical technology field. Using machine learning methods, a data-driven approach can be adopted to accurately predict and optimize the uniformity of cartilage powder tableting, thereby improving production efficiency and product quality.

[0003] While these methods can provide a certain level of control accuracy, they mostly rely on a single data source and traditional linear modeling methods, making it difficult to effectively handle complex nonlinear relationships and high-dimensional data. When dealing with complex production environments and integrating multi-dimensional information, these technologies often face problems such as insufficient prediction accuracy and slow response speed. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a machine learning-based method for predicting the uniformity of cartilage powder tablet content, which solves the problem of not being able to effectively handle complex nonlinear relationships and high-dimensional data.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] This invention provides a machine learning-based method for predicting the uniformity of cartilage powder tablet content. The method includes: collecting multi-dimensional cartilage powder tablet data and preprocessing the data to generate a tablet feature dataset; using a particle swarm optimization algorithm to select features and fuse data from the tablet feature dataset to obtain a selected feature set, extracting process features from the selected feature set, and combining these process features with environmental and production parameters to generate a training feature dataset; using an L1 regularization algorithm to remove redundant features from the training feature dataset, inputting the training feature dataset into a deep neural network, and performing nonlinear mapping on the redundant features to generate a uniformity prediction value; using a difference method to backtrack the time-series error of the uniformity prediction value to obtain the uniformity deviation, and adjusting the pressing force and temperature / humidity conditions within the uniformity deviation to generate a uniformity error range; and based on the uniformity error range, adjusting the uniformity deviation of the production parameters through a reward mechanism, and optimizing the pressing force and temperature / humidity process conditions to generate an optimized control scheme.

[0008] As a preferred embodiment of the machine learning-based method for predicting the uniformity of cartilage powder tablet content described in this invention, the specific steps for collecting multi-dimensional cartilage powder tablet data and preprocessing the multi-dimensional cartilage powder tablet data to generate a tablet feature dataset are as follows:

[0009] Collect multi-dimensional cartilage powder tablet data, denoise the multi-dimensional cartilage powder tablet data, obtain tablet data, fill missing values ​​and remove abnormal data points to generate cleaned tablet data;

[0010] Multi-dimensional features are extracted from the tableting data after cleaning, and the shape contour and defect edges of the tableting surface are identified by edge detection method to generate a tableting feature dataset.

[0011] As a preferred embodiment of the machine learning-based method for predicting the uniformity of cartilage powder tablet content described in this invention, the step of using a particle swarm optimization algorithm to perform feature selection and data fusion on the tablet feature dataset to obtain a selected feature set includes the following specific steps:

[0012] The particle swarm optimization algorithm is used to select features from the tablet feature dataset. By simulating the movement of particles in the solution space, the current position and current velocity of each particle are calculated to generate a candidate feature set.

[0013] Time alignment and information integration are performed on the tablet compression feature dataset to obtain a comprehensive feature set. The feature contribution of the candidate feature set is prioritized to generate a selected feature set.

[0014] As a preferred embodiment of the machine learning-based method for predicting the uniformity of cartilage powder tablet content described in this invention, the specific steps for extracting process features from a selected feature set and combining these process features with environmental and production parameters to generate a training feature dataset are as follows:

[0015] The correlation between each process feature in the selected feature set and the uniformity prediction is analyzed to form a process feature set;

[0016] The process feature set is combined with environmental and production parameters, and a training feature dataset is generated through numerical standardization and timestamp alignment.

[0017] As a preferred embodiment of the machine learning-based method for predicting the uniformity of cartilage powder tablet content described in this invention, the steps of removing redundant features from the training feature dataset using the L1 regularization algorithm to obtain a simplified feature set, inputting the simplified feature set into a deep neural network, and performing a nonlinear mapping on the simplified feature set to generate a uniformity prediction value are as follows:

[0018] The L1 regularization algorithm is used to apply a penalty term to the training feature dataset, and feature weight compression is performed to remove redundant features from the training feature dataset, generating a simplified feature set.

[0019] The simplified feature set is input into a deep neural network, and a non-linear transformation is performed on the simplified feature set to generate an intermediate feature representation;

[0020] Feature mapping is performed on the intermediate feature representations to adjust the dependencies between the simplified feature sets and generate uniformity prediction values.

[0021] As a preferred embodiment of the machine learning-based method for predicting the uniformity of cartilage powder tablet content described in this invention, the specific steps for performing feature mapping on intermediate feature representations, adjusting the dependencies between simplified feature sets, and generating uniformity prediction values ​​are as follows:

[0022] The intermediate feature representations are nonlinearly transformed using the t-SNE method, the similarity between intermediate feature representations is calculated, and the feature point layout in the low-dimensional space is optimized using the gradient descent method, which is then mapped to the low-dimensional space to generate optimized feature representations.

[0023] The optimized feature representations are numerically normalized, and the dependencies between the optimized feature representations are adjusted to generate uniformity prediction values.

[0024] As a preferred embodiment of the machine learning-based method for predicting the uniformity of cartilage powder tablet content described in this invention, the step of obtaining the uniformity deviation by performing time-series error backtracking on the predicted uniformity value using the difference method is as follows:

[0025] The difference between the current uniformity prediction value and the uniformity prediction value at the previous time step is calculated using the finite difference method to generate the uniformity time series error.

[0026] By combining the time-series error of uniformity with the historical error backtracking, error adjustment is performed to generate uniformity deviation.

[0027] As a preferred embodiment of the machine learning-based method for predicting the uniformity of cartilage powder tablet content described in this invention, the uniformity error range is obtained by calculating the adjustment range based on the uniformity deviation and by making slight adjustments to the pressing force and temperature and humidity conditions through a multiple linear regression method.

[0028] As a preferred embodiment of the machine learning-based method for predicting the uniformity of cartilage powder tablet content described in this invention, the steps of adjusting the uniformity deviation of production parameters through a reward mechanism based on the uniformity error range, and optimizing the pressing force and temperature and humidity process conditions to generate an optimized control scheme are as follows:

[0029] The uniformity error range is scaled and nonlinearly compressed to obtain the reward adjustment coefficient. Based on the reward adjustment coefficient, a stable interval constraint analysis is performed on the current combination of pressing force and temperature and humidity to generate the process adjustment intensity value.

[0030] Based on the process adjustment intensity value, the pressing force and temperature and humidity process conditions of the process feature concentration are simultaneously and slightly adjusted to adjust the uniformity deviation in the production process, optimize the pressing force and temperature and humidity process conditions, and generate an optimized control scheme.

[0031] As a preferred embodiment of the machine learning-based method for predicting the uniformity of cartilage powder tablet content described in this invention, the step of simultaneously and slightly adjusting the pressing force and temperature / humidity process conditions within the process feature set based on the process adjustment intensity value, adjusting the uniformity deviation during production, optimizing the pressing force and temperature / humidity process conditions, and generating an optimized control scheme, is as follows:

[0032] The adjustment level is determined based on the process adjustment intensity value, and the adjustment range is selected according to the adjustment level to generate an adjustment strategy;

[0033] Based on the adjustment strategy, the records of pressing force and temperature and humidity disturbances of adjacent batches are called to construct an error response sequence and generate a response slope vector.

[0034] The synchronous adjustment step size is calculated using the response slope vector and the process adjustment intensity value, and micro-amplitude synchronous adjustment is performed to generate an optimized control scheme.

[0035] The beneficial effects of this invention are as follows: by collecting and preprocessing multi-dimensional cartilage powder tableting data, a high-quality tableting feature dataset is generated, ensuring the accuracy and consistency of the data; by using particle swarm optimization algorithm and machine learning method to select and fuse feature data, a carefully selected feature set is obtained, thus optimizing the feature selection process; through these two steps, the accuracy and computational efficiency of uniformity prediction are improved, providing reliable machine learning method support for the optimization of the cartilage powder tableting process. Attached Figure Description

[0036] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the 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.

[0037] Figure 1 This is a flowchart of a machine learning-based method for predicting the uniformity of cartilage powder content in tablets.

[0038] Figure 2 This is a flowchart of a method for predicting the uniformity of cartilage powder content in tablets.

[0039] Figure 3 A flowchart for generating the tablet compression feature dataset.

[0040] Figure 4 A flowchart generated to optimize the control scheme. Detailed Implementation

[0041] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0042] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0043] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0044] Reference Figures 1-4 This is one embodiment of the present invention, which provides a machine learning-based method for predicting the uniformity of cartilage powder content in tablets, comprising the following steps:

[0045] S1. Collect multi-dimensional cartilage powder tablet data and preprocess the multi-dimensional cartilage powder tablet data to generate a tablet feature dataset.

[0046] S1.1 Collect multi-dimensional cartilage powder tablet data, perform noise reduction on the multi-dimensional cartilage powder tablet data, obtain tablet data, fill in missing values ​​and remove abnormal data points to generate cleaned tablet data.

[0047] It should be noted that real-time data acquisition is performed at different locations using multiple high-precision sensor arrays, covering key parameters during the tableting process, such as compression pressure, temperature, humidity, and density. The acquired tableting data then undergoes a denoising process. During denoising, wavelet transform algorithms are used to suppress noise in the signal, preserving meaningful signal components to ensure data accuracy and stability. Missing values ​​are filled using interpolation to ensure reasonable estimations at the locations of missing data, avoiding impacts on subsequent analysis due to incomplete data. For outlier data points, statistical analysis methods are used to identify and remove abnormal values ​​exceeding the normal range, generating cleaned tableting data.

[0048] S1.2 Extract multi-dimensional features from the tableting data after cleaning, and identify the shape contour and defect edges of the tableting surface through edge detection method to generate a tableting feature dataset.

[0049] It should be noted that feature extraction is performed on the tablet compression data, including multi-dimensional features such as tablet thickness, density, and uniformity. Each multi-dimensional feature is quantified using statistical analysis methods such as mean and standard deviation to obtain a comprehensive description of the tablet compression process. When extracting shape contours and defect edges, edge detection methods are used to analyze the geometry of the tablet surface. Specific steps include applying edge detection algorithms and using gradient calculations to identify minute edge variations and defect locations on the tablet surface. The edge images are further optimized through morphological operations such as dilation and erosion to ensure the clarity of defect edges, generating a tablet compression feature dataset.

[0050] S2. The particle swarm optimization algorithm is used to select features and fuse data in the tableting feature dataset to obtain a selected feature set. Process features are then extracted from the selected feature set and combined with environmental and production parameters to generate a training feature dataset.

[0051] It should be noted that existing methods employ traditional feature selection techniques, such as filtering, wrapping, or embedding methods, which rely on simple statistical or correlation analysis to select features. Data fusion generally relies on pre-defined rules or simple merging methods to combine data from different sources, potentially ignoring nonlinear relationships and complex interactions between data. In feature extraction, existing methods may primarily focus on single-dimensional process features or environmental parameters, lacking comprehensive consideration of multi-dimensional data.

[0052] This invention employs a particle swarm optimization algorithm for feature selection and data fusion. By intelligently optimizing and selecting the features most relevant to tablet uniformity prediction, it maximizes the predictive power of the feature set. Simultaneously, it combines process features with environmental and production parameters, using comprehensive data from these parameters to generate a training feature dataset. This more accurately reflects the interactions and influences between multi-dimensional data. This method better captures the complex relationships between data, improving the accuracy and reliability of predictions.

[0053] S2.1. The particle swarm optimization algorithm is used to select features from the tablet feature dataset. By simulating the movement of particles in the solution space, the current position and current velocity of each particle are calculated to generate a candidate feature set.

[0054] It should be noted that each tablet compression feature is considered as a particle in a particle swarm, and each particle is assigned an initial position and velocity. The particle's current position represents the feature's selection state in the feature space, while the particle's velocity determines the step size of its movement in the solution space. The Particle Swarm Optimization (PSO) algorithm searches for the optimal subset of features by simulating the movement of particles in the solution space. In each iteration, the particle adjusts its velocity and position based on the prediction accuracy evaluation of the selected feature set. The PSO algorithm finds the optimal feature set by maximizing prediction accuracy, ensuring that the selected features can most effectively predict tablet uniformity. In each iteration, the particle's velocity and position are updated based on the guidance of local and global optima until the PSO algorithm converges, generating a candidate feature set.

[0055] S2.2 Perform time alignment and information integration on the tablet compression feature dataset to obtain a comprehensive feature set, and prioritize the feature contribution of the candidate feature set to generate a selected feature set.

[0056] It should be noted that each feature is synchronized according to the data acquisition timestamp to ensure that data with different features can be compared on the same time scale. This includes aligning data from different time points using interpolation methods to ensure temporal consistency between data for effective analysis. Through information integration, multi-dimensional feature data are combined to form a comprehensive feature set. During the integration process, different process features are standardized to eliminate the influence of dimensional differences, making the values ​​of each feature comparable. The candidate feature set and the comprehensive feature set are prioritized. The ranking is based on the correlation between each feature and uniformity prediction and its contribution to prediction accuracy. The ranking is achieved by calculating the importance index of each feature in tablet uniformity prediction, generating a carefully selected feature set.

[0057] S2.3 Analyze the correlation between each process feature in the selected feature set and the uniformity prediction to form a process feature set.

[0058] It should be noted that the linear relationship between each process feature and the predicted uniformity value is quantified by calculating the Pearson correlation coefficient. The Pearson correlation coefficient ranges from -1 to +1, where +1 indicates a perfect positive correlation, -1 indicates a perfect negative correlation, and 0 indicates no correlation. By calculating the correlation coefficient between each process feature and the predicted uniformity value, process features highly correlated with the predicted uniformity value are selected, while features with low or redundant correlation coefficients are excluded. The strong correlation between these process features and the predicted uniformity value reflects their influence on tablet uniformity. The selected process features are used as a process feature set, and further multivariate regression analysis is used to evaluate the independent contribution of each process feature to uniformity prediction. Finally, after analysis, the process features with the greatest influence on uniformity prediction are selected to form the process feature set.

[0059] It should also be noted that the Pearson correlation coefficient is a statistical indicator used to measure the strength and direction of the linear relationship between two variables. Its value ranges from -1 to +1, where +1 indicates a perfectly positive linear correlation, -1 indicates a perfectly negative linear correlation, and 0 indicates no linear correlation.

[0060] S2.4 Combine the process feature set with environmental and production parameters, and generate a training feature dataset through numerical standardization and timestamp alignment.

[0061] It should be noted that aligning each feature in the process feature set with its corresponding environmental and production parameters is accomplished through timestamp alignment, ensuring accurate correspondence between data at the same point in time. For example, process parameters such as pressing force, temperature, and humidity are respectively mapped one-to-one with environmental parameters such as temperature and humidity, as well as operational parameters during production. During alignment, interpolation is used to fill in timestamp mismatches caused by differences in collection frequency, ensuring all data points have the same time scale. Subsequently, numerical standardization unifies the numerical range of each parameter. For example, Z-score standardization adjusts the mean of all process features to 0 and the standard deviation to 1, thereby eliminating the impact of differences in feature dimensions on data analysis. After standardization and timestamp alignment, all feature data are integrated to generate a training feature dataset.

[0062] S3. Redundant features in the training feature dataset are removed by L1 regularization algorithm. The training feature dataset is then input into a deep neural network, and the redundant features are nonlinearly mapped to generate uniformity prediction values.

[0063] S3.1. Apply a penalty term to the training feature dataset using the L1 regularization algorithm, and perform feature weight compression to remove redundant features from the training feature dataset, generating a simplified feature set.

[0064] It should be noted that a penalty term is introduced into the weight of each process feature, which is the sum of the absolute values ​​of each feature weight. The purpose of L1 regularization is to select features by increasing the penalty term so that the weights of some redundant features approach zero. During training, the L1 regularization algorithm continuously updates the weight of each feature and compresses and removes redundant features from the training feature dataset. When the weight of a feature is close to zero, the redundant feature is considered to have little impact on uniformity prediction and is therefore removed from the training feature dataset. Through this process, the L1 regularization algorithm can effectively reduce redundant features in the feature set, retaining only those features that affect uniformity prediction, generating a simplified feature set.

[0065] It should also be noted that L1 regularization is a method used for feature selection and preventing overfitting, which achieves this by penalizing the model's weights. L1 regularization adds the sum of the absolute values ​​of all feature weights as a penalty term to the loss function. This means that during optimization, the model must not only minimize the prediction error but also minimize the feature weights as much as possible.

[0066] The expression for calculating the sum of the absolute values ​​of each feature weight is:

[0067] ;

[0068] in, The sum of the absolute values ​​of the weights of each process feature. This is a regularization parameter that controls the strength of regularization. For the first The weight of each process feature; The total number of features.

[0069] S3.2 Input the simplified feature set into the deep neural network, perform nonlinear transformation on the simplified feature set, and generate intermediate feature representation.

[0070] It should be noted that the simplified feature set undergoes preprocessing to ensure it meets the input requirements of deep neural networks, such as standardization and outlier removal. Deep neural networks include various network architectures, such as multilayer perceptrons, convolutional neural networks, and recurrent neural networks. The processed simplified feature set is passed through the input layer of the multilayer perceptron to the first hidden layer. In each layer, the multilayer perceptron performs nonlinear transformations on the input simplified feature set, progressively enhancing its ability to capture complex relationships between features. The output of each layer serves as the input to the next layer, and features are continuously transformed and combined between layers, gradually extracting higher-level abstract features. In this process, the multilayer perceptron transforms the information in the simplified feature set into intermediate feature representations through multiple nonlinear transformations.

[0071] It should also be noted that the training process of a Multilayer Perceptron (MLP) involves forward propagation, passing the input data layer by layer to the output layer to generate prediction results. Then, the error between the prediction results and the true values ​​is calculated, and the gradient of the error is calculated layer by layer using the backpropagation algorithm to adjust the weights and biases. Through multiple iterations of training, the loss is minimized, thereby improving the model's prediction accuracy.

[0072] S3.3. The intermediate feature representations are nonlinearly transformed using the t-SNE method, the similarity between intermediate feature representations is calculated, and the feature point layout in the low-dimensional space is optimized using the gradient descent method, which is then mapped to the low-dimensional space to generate optimized feature representations.

[0073] It should be noted that the t-SNE method typically uses Euclidean distance to measure the similarity between intermediate feature representations. The t-SNE method maps high-dimensional feature representations to a low-dimensional space, aiming to maximize the proximity of similar features in the low-dimensional space while distancing dissimilar features. t-SNE optimizes the feature point distribution in the low-dimensional space using gradient descent, continuously adjusting the position of each feature point to bring similar features closer together and separate dissimilar features. The optimization process gradually improves the distribution of feature points in the low-dimensional space by minimizing the error, generating optimized feature representations.

[0074] It should also be noted that the t-SNE method is a non-linear dimensionality reduction method, mainly used for the visualization of high-dimensional data. Its core idea is to map data points in high-dimensional space to low-dimensional space as much as possible by preserving the similarity between data points, while maintaining the local structure of the data.

[0075] S3.4 Perform numerical normalization on the optimized feature representations and adjust the dependencies between the optimized feature representations to generate uniformity prediction values.

[0076] It should be noted that the values ​​of each process feature are uniformly scaled to convert them into a common numerical range. Common normalization methods include dividing the feature values ​​by the difference between the mean and the standard deviation, ensuring that the mean of each process feature is 0 and the standard deviation is 1, thus ensuring that dimensional differences between different features do not affect subsequent analysis. Adjusting and optimizing the dependencies between feature representations mainly involves analyzing the correlation between features, removing redundant features, and retaining features with independent information. This step typically uses principal component analysis to assess the dependencies between features. By adjusting the relationships between features, it is ensured that the final optimized feature representation effectively reflects the key factors in uniformity prediction, generating uniformity prediction values.

[0077] S4. The time-series error of the uniformity prediction value is backtracked by the difference method to obtain the uniformity deviation, and the pressure and temperature and humidity conditions in the uniformity deviation are adjusted to generate the uniformity error range.

[0078] S4.1 Calculate the difference between the current uniformity prediction value and the uniformity prediction value at the previous time step using the difference method to generate the uniformity time series error.

[0079] It should be noted that the uniformity prediction value at the current time step and the uniformity prediction value at the previous time step are calculated using the finite difference method. The error value between these two uniformity prediction values ​​reflects the magnitude of change between the current prediction result and the previous prediction result. The key to the finite difference method is calculating the amount of change between two time points; therefore, it is necessary to compare the uniformity prediction values ​​at the two time points in each iteration. The uniformity time series error is obtained through calculation.

[0080] It should also be noted that the finite difference method is a technique used to analyze changes in time series data. By calculating the difference between adjacent data points, it reveals the trend of data change. In uniformity forecasting, the finite difference method helps determine the difference between the current forecast value and the forecast value at the previous time point, thereby quantifying the change in uniformity forecast. This method can effectively extract time series errors, providing a basis for subsequent error adjustment.

[0081] S4.2 Combine the time-series error of uniformity with the historical error backtracking to perform error adjustment and generate uniformity deviation.

[0082] It should be noted that the uniformity time-series error reflects the variation degree of the predicted uniformity values between different time points. The current uniformity time-series error is compared with historical errors to evaluate the change direction and amplitude of the error. If the change trend of the current error is consistent with that of the historical errors, it may be an error continuation, which needs to be corrected by increasing the adjustment intensity; if the trends are opposite, it indicates that the prediction is improved, and the adjustment amplitude can be appropriately reduced. Based on the rules in the historical data and the specific situation of the current error, and combined with historical errors, the current error is adaptively adjusted to optimize the accuracy of uniformity prediction. This method can effectively eliminate the influence of accidental errors, ensure the stability and accuracy of the predicted uniformity values, and generate a uniformity deviation.

[0083] S4.3. Calculate the adjustment amplitude according to the uniformity deviation, and make slight adjustments to the pressing force as well as temperature and humidity conditions through a multiple linear regression method to generate a uniformity error range.

[0084] It should be noted that the uniformity deviation value of each tablet pressing sample is calculated, and this difference value represents the deviation degree between the target uniformity and the predicted uniformity. The uniformity deviations of all samples are summarized, and the adjustment amplitude of the overall uniformity deviation is obtained through a statistical method, that is, how to fine-tune the tablet pressing process parameters to reduce errors. The multiple linear regression method is used to make slight adjustments to the pressing force as well as temperature and humidity conditions. By taking the uniformity deviation as the dependent variable and the pressing force as well as temperature and humidity conditions as the independent variables, regression analysis is carried out to calculate the contribution degree and adjustment direction of each independent variable to the uniformity deviation, so as to generate the uniformity error range.

[0085] S5. Adjust the uniformity deviation of production parameters through a reward mechanism according to the uniformity error range, and optimize the process conditions of pressing force, temperature and humidity to generate an optimized control scheme.

[0086] It should be noted that existing methods control the uniformity deviation by setting fixed process parameter ranges (such as pressing force, temperature and humidity conditions), and the adjustment process relies on manual experience or a simple rule engine. These methods may adopt preset standardized adjustment strategies, lack precise response to dynamic changes in actual production, and the adjustment process is often not adaptive, so it may not be possible to optimize parameters in a timely manner for complex production environments.

[0087] The present invention dynamically adjusts the uniformity deviation of production parameters based on the uniformity error range by introducing a reward mechanism. This method combines real-time feedback and an optimization algorithm to automatically identify deviations and perform precise adjustment, which can not only optimize the process conditions of pressing force, temperature and humidity, but also has adaptive adjustment capability, so as to improve the stability and uniformity of the process in a constantly changing production environment. This intelligent adjustment mechanism can more effectively cope with various changes in the production process and optimize the control scheme in real time.

[0088] S5.1. Obtain the reward adjustment coefficient by scaling and nonlinear compression of the uniformity error range, and perform stable interval constraint analysis on the current combination of pressing force and temperature and humidity based on the reward adjustment coefficient to generate the process adjustment intensity value.

[0089] It should be noted that all data within the uniformity error range are normalized to ensure consistent dimensions, enabling comparisons on the same scale. Nonlinear compression is then applied to the error range data, reducing the impact of extreme values ​​and smoothing the data even under large variations, thus preventing outliers from significantly affecting the results. The ultimate goal of this process is to generate a reward adjustment coefficient, which reflects the impact of the uniformity error range on the adjustment strategy. Based on this coefficient, the adjustment ranges for pressing force and temperature / humidity are determined, generating process adjustment intensity values.

[0090] S5.2 Determine the adjustment level based on the process adjustment intensity value, select the adjustment range according to the adjustment level, and generate the adjustment strategy.

[0091] It should be noted that, based on the magnitude of the process adjustment intensity value, the adjustment process is divided into multiple levels according to the standard. A large adjustment intensity value indicates a significant uniformity error, and a stronger demand for adjustment of the tableting process; cases with large process adjustment intensity values ​​are classified as high adjustment levels. Conversely, small adjustment intensity values ​​correspond to low adjustment levels. Each adjustment level corresponds to a different adjustment range, the selection of which is determined based on the process adjustment intensity value and the tolerance for uniformity error. High adjustment levels will choose a large adjustment range to ensure sufficient adjustment of process conditions, while low adjustment levels will choose a small adjustment range to avoid over-adjustment. After determining the adjustment level and adjustment range, an adjustment strategy is generated.

[0092] S5.3. Based on the adjustment strategy, call the records of pressing force and temperature and humidity disturbances of adjacent batches, construct the error response sequence, and generate the response slope vector.

[0093] It should be noted that accessing the records of pressing force and temperature / humidity disturbances from adjacent batches reflects the changes in pressing force and temperature / humidity between batches. After acquiring these disturbance records, an error response sequence is constructed by comparing them with the uniformity deviation data of the corresponding batches. The error response sequence correlates each disturbance change with the change in uniformity deviation, reflecting the impact of different pressing force and temperature / humidity adjustments on uniformity error. A response slope vector is generated by calculating the rate of change of each data point in the error response sequence.

[0094] S5.4 Calculate the synchronous adjustment step size using the response slope vector and process adjustment intensity value, perform micro-amplitude synchronous adjustment, and generate an optimized control scheme.

[0095] It should be noted that the adjustment range for each process parameter is determined based on the direction and magnitude of the response slope vector. The response slope vector provides information on the sensitivity of each parameter to uniformity error; a large slope indicates that adjusting the parameter has a significant impact on uniformity error, requiring a larger adjustment. The micro-amplitude synchronous adjustment step size is calculated by combining the process adjustment intensity value with the requirement for uniformity deviation. The adjustment step size represents the magnitude of the process parameter change during each adjustment, ensuring that the adjustment process is not too drastic, thus maintaining process stability. Micro-amplitude synchronous adjustment means that in each iteration, the pressing force and temperature / humidity will be adjusted synchronously and subtly to ensure that the uniformity error converges within a reasonable range, generating an optimized control scheme.

[0096] In summary, this invention achieves the following: by collecting and preprocessing multi-dimensional cartilage powder tableting data to generate a high-quality tableting feature dataset, ensuring data accuracy and consistency; and by employing particle swarm optimization and machine learning methods to select and fuse feature data, obtaining a carefully selected feature set, thus optimizing the feature selection process. Through these two steps, the accuracy and computational efficiency of uniformity prediction are improved, providing reliable machine learning support for optimizing the cartilage powder tableting process.

[0097] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for predicting the uniformity of cartilage powder content in tablets based on machine learning, characterized in that, include: Collect multi-dimensional cartilage powder tablet data and preprocess the multi-dimensional cartilage powder tablet data to generate a tablet feature dataset; The particle swarm optimization algorithm is used to select features and fuse data in the tableting feature dataset to obtain a selected feature set. Process features are then extracted from the selected feature set and combined with environmental and production parameters to generate a training feature dataset. The L1 regularization algorithm is used to remove redundant features from the training feature dataset to obtain a simplified feature set. The simplified feature set is then input into a deep neural network, and a nonlinear mapping is performed on the simplified feature set to generate a uniformity prediction value. By using the finite difference method to backtrack the time-series error of the uniformity prediction value, the uniformity deviation is obtained, and the pressure and temperature and humidity conditions in the uniformity deviation are adjusted to generate the uniformity error range. Based on the uniformity error range, the uniformity deviation of production parameters is adjusted through a reward mechanism, and the pressing force and temperature and humidity process conditions are optimized to generate an optimized control scheme.

2. The method for predicting the uniformity of cartilage powder content in tablets based on machine learning as described in claim 1, characterized in that, The process of collecting multi-dimensional cartilage powder tablet data and preprocessing the multi-dimensional cartilage powder tablet data to generate a tablet feature dataset involves the following steps: Collect multi-dimensional cartilage powder tablet data, denoise the multi-dimensional cartilage powder tablet data, obtain tablet data, fill missing values ​​and remove abnormal data points to generate cleaned tablet data; Multi-dimensional features are extracted from the tableting data after cleaning, and the shape contour and defect edges of the tableting surface are identified by edge detection method to generate a tableting feature dataset.

3. The method for predicting the uniformity of cartilage powder content in tablets based on machine learning as described in claim 2, characterized in that, The particle swarm optimization algorithm is used to perform feature selection and data fusion on the tablet compression feature dataset to obtain a carefully selected feature set. The specific steps are as follows: The particle swarm optimization algorithm is used to select features from the tablet feature dataset. By simulating the movement of particles in the solution space, the current position and current velocity of each particle are calculated to generate a candidate feature set. Time alignment and information integration are performed on the tablet compression feature dataset to obtain a comprehensive feature set. The feature contribution of the candidate feature set is prioritized to generate a selected feature set.

4. The method for predicting the uniformity of cartilage powder content in tablets based on machine learning as described in claim 3, characterized in that, The specific steps for extracting process features from the selected feature set and combining these features with environmental and production parameters to generate a training feature dataset are as follows: The correlation between each process feature in the selected feature set and the uniformity prediction is analyzed to form a process feature set; The process feature set is combined with environmental and production parameters, and a training feature dataset is generated through numerical standardization and timestamp alignment.

5. The method for predicting the uniformity of cartilage powder content in tablets based on machine learning as described in claim 4, characterized in that, The process involves using the L1 regularization algorithm to remove redundant features from the training feature dataset, obtaining a simplified feature set, and then inputting the simplified feature set into a deep neural network. A non-linear mapping is then performed on the simplified feature set to generate uniformity prediction values. The specific steps are as follows: The L1 regularization algorithm is used to apply a penalty term to the training feature dataset, and feature weight compression is performed to remove redundant features from the training feature dataset, generating a simplified feature set. The simplified feature set is input into a deep neural network, and a non-linear transformation is performed on the simplified feature set to generate an intermediate feature representation; Feature mapping is performed on the intermediate feature representations to adjust the dependencies between the simplified feature sets and generate uniformity prediction values.

6. The method for predicting the uniformity of cartilage powder content in tablets based on machine learning as described in claim 5, characterized in that, The specific steps for performing feature mapping on the intermediate feature representations, adjusting the dependencies between the simplified feature sets, and generating uniformity prediction values ​​are as follows: The intermediate feature representations are nonlinearly transformed using the t-SNE method, the similarity between intermediate feature representations is calculated, and the feature point layout in the low-dimensional space is optimized using the gradient descent method, which is then mapped to the low-dimensional space to generate optimized feature representations. The optimized feature representations are numerically normalized, and the dependencies between the optimized feature representations are adjusted to generate uniformity prediction values.

7. The method for predicting the uniformity of cartilage powder content in tablets based on machine learning as described in claim 1, characterized in that, The process of using the difference method to backtrack the time-series error of the uniformity prediction value and obtain the uniformity deviation is as follows: The difference between the current uniformity prediction value and the uniformity prediction value at the previous time step is calculated using the finite difference method to generate the uniformity time series error. By combining the time-series error of uniformity with the historical error backtracking, error adjustment is performed to generate uniformity deviation.

8. The method for predicting the uniformity of cartilage powder content in tablets based on machine learning as described in claim 1, characterized in that, The uniformity error range is calculated based on the uniformity deviation and adjusted slightly by using a multiple linear regression method to adjust the pressing force and temperature and humidity conditions.

9. The method for predicting the uniformity of cartilage powder content in tablets based on machine learning as described in claim 8, characterized in that, The process involves adjusting the uniformity deviation of production parameters based on the uniformity error range using a reward mechanism, optimizing the pressing force and temperature / humidity process conditions, and generating an optimized control scheme. The specific steps are as follows: The uniformity error range is scaled and nonlinearly compressed to obtain the reward adjustment coefficient. Based on the reward adjustment coefficient, a stable interval constraint analysis is performed on the current combination of pressing force and temperature and humidity to generate the process adjustment intensity value. Based on the process adjustment intensity value, the pressing force and temperature and humidity process conditions of the process feature concentration are simultaneously and slightly adjusted to adjust the uniformity deviation in the production process, optimize the pressing force and temperature and humidity process conditions, and generate an optimized control scheme.

10. The method for predicting the uniformity of cartilage powder content in tablets based on machine learning as described in claim 9, characterized in that, The process involves synchronously and slightly adjusting the pressing force and temperature / humidity conditions within the process feature set based on the process adjustment intensity value. This adjusts the uniformity deviation during production, optimizes the pressing force and temperature / humidity conditions, and generates an optimized control scheme. The specific steps are as follows: The adjustment level is determined based on the process adjustment intensity value, and the adjustment range is selected according to the adjustment level to generate an adjustment strategy; Based on the adjustment strategy, the records of pressing force and temperature and humidity disturbances of adjacent batches are called to construct an error response sequence and generate a response slope vector. The synchronous adjustment step size is calculated using the response slope vector and the process adjustment intensity value, and micro-amplitude synchronous adjustment is performed to generate an optimized control scheme.

Images