An intelligent dropwise adding control system for medicine production

Abstract

This invention provides an intelligent droplet addition control system for pharmaceutical production, comprising a basic information acquisition unit, a droplet shape prediction unit, a real-time shape acquisition unit, a shape difference recognition unit, and a droplet addition control unit. The basic information acquisition unit collects and integrates forming features and sends them to the droplet shape prediction unit. The droplet shape prediction unit processes the forming parameters based on a built-in deep learning network and generates a standard shape image. The real-time shape acquisition unit can acquire real-time shape images of the droplets. The shape difference recognition unit identifies the differences between the real-time shape image and the standard shape image and generates a difference analysis report. The droplet addition control unit generates a control strategy based on the difference analysis report and controls the production equipment. By comparing and identifying the droplets, it can determine whether there are any abnormalities in the production equipment, thereby adjusting and controlling the production equipment to ensure the generation of standard droplets and guarantee the quality of pharmaceutical production.

Description

Technical Field

[0001] This invention relates to the field of pharmaceutical processing technology, and in particular to an intelligent dripping control system for pharmaceutical production. Background Technology

[0002] In the pharmaceutical manufacturing industry, raw material dropwise addition is a key process that runs through many core links such as chemical synthesis and pharmaceutical manufacturing, preparation of biological agents, and extraction and purification of traditional Chinese medicine. For example, the precise synthesis of anti-tumor drugs, the addition of excipients in vaccine production, and the gradient formulation of traditional Chinese medicine extracts all rely on stable dropwise addition operations. The control precision of the dropwise addition process directly determines the dosage consistency, efficacy stability and safety of the drug, and is one of the core prerequisites for ensuring that the drug meets the standards.

[0003] The physicochemical properties of different pharmaceutical raw materials vary significantly, encompassing low-viscosity solvents, high-viscosity concentrates, suspensions, and special solutions containing surfactants. This necessitates a high degree of adaptability to the dropping process. Currently, commonly used dropping equipment often employs single flow or pressure parameters for control, focusing only on macroscopic flow data and neglecting the correlation between droplet shape and raw material characteristics and the state of the dropping head. This makes it impossible to identify droplet formation anomalies caused by fluctuations in raw material concentration, viscosity changes, or dropping head wear / crystallization. Furthermore, the lack of adaptability to different types of raw materials makes it difficult to match the dropping requirements of diverse pharmaceutical solutions, easily leading to problems such as dosage deviation, satellite droplets, and stringing. In addition, the control of the dropping process is mostly reactive, unable to predict anomalies in real time and provide precise control. This results in large batch quality fluctuations, significant waste of valuable raw materials, and an inability to meet the stringent requirements for dropping precision in high-end biological agents and precision medicine drugs, thus hindering the intelligent and high-quality development of pharmaceutical production. Summary of the Invention

[0004] In view of this, the present invention proposes an intelligent dripping control system for pharmaceutical production, which can compare the real-time dripping droplets with standard droplets, generate a dripping control strategy based on the comparison results, adjust the dripping process, and ensure the quality of pharmaceutical production.

[0005] The technical solution of this invention is implemented as follows: An intelligent dripping control system for pharmaceutical production includes: The basic information acquisition unit is used to collect data on the type of raw material to be added, raw material concentration, required dripping flow rate, dripping head model, and dripping environment, and integrate them into forming features; The droplet shape prediction unit is used to process the forming parameters using a pre-built deep learning network to obtain a standard shape image of the raw material droplet; The real-time shape acquisition unit is used to acquire real-time shape images of the droplets formed by the raw materials being added by the dispensing head; The shape difference recognition unit is used to perform difference recognition between standard shape images and real-time shape images, and obtain a difference analysis report between the real-time shape image and the standard shape image; The dispensing control unit is used to determine the control strategy based on the difference analysis report and to control the dispensing head and / or concentration adjustment mechanism according to the control strategy; The basic information acquisition unit is connected to the drop shape prediction unit, and the shape difference recognition unit is connected to the drop shape prediction unit, the real-time shape acquisition unit, and the drop control unit, respectively.

[0006] Preferably, the execution steps of the basic information collection unit include: Obtain the drug production requirements, and extract the types of raw materials, concentrations of raw materials, required dripping flow rates, and models of dripping heads from these requirements. Temperature, humidity, and air pressure data of the environment where the dispensing head is located are collected by the set temperature and humidity sensors and air pressure sensors; Feature extraction was performed on raw material type, raw material concentration, required dripping flow rate, dripping head model, temperature and humidity data, and air pressure data, respectively. The extracted features are concatenated into vector features, and the vector features are output as shaped features.

[0007] Preferably, the execution steps of the droplet shape prediction unit include: Receive and parse the formed features transmitted from the basic information acquisition unit; The parsed shaped features are then input into a pre-constructed convolutional neural network; The shaping parameters are subjected to multi-layer nonlinear transformation and feature extraction through the forward propagation process of a convolutional neural network; The output layer of the convolutional neural network outputs standard image data containing the standard shape of a droplet.

[0008] Preferably, it further includes a forming time determination unit, used to determine the forming time of the droplet based on the forming characteristics, and to input the forming time to the real-time shape acquisition unit. The forming time determination unit is connected to the basic information acquisition unit and the real-time shape acquisition unit respectively.

[0009] Preferably, the execution steps of the real-time shape acquisition unit include: A high-speed camera is deployed near the dripping head to collect image data of the entire process of raw material dripping from the dripping head. The entire process image data is decomposed frame by frame to obtain the starting frame image; Based on the forming time, the forming frame image is obtained by shifting backward from the starting frame image, and the forming frame image is output as a real-time shape image of the droplet.

[0010] Preferably, it also includes an image processing unit for performing noise reduction, grayscale conversion, and edge enhancement processing on the standard shape image and the real-time shape image. The image processing unit is connected to the droplet shape prediction unit, the real-time shape acquisition unit, and the shape difference recognition unit, respectively.

[0011] Preferably, the execution steps of the shape difference recognition unit include: The SIFT algorithm was used to extract the contour feature points of the droplets in real-time shape images and standard shape images; Two sets of contour feature points are initially matched by a fast approximate nearest neighbor search, and mismatched points after the initial matching are removed to obtain accurately matched feature pairs. The precisely matched image is input into a lightweight CNN model, and the standardized cross-correlation coefficient is calculated on the feature map output by the lightweight CNN model. The standardized cross-correlation coefficient is compared with a preset threshold to identify and mark the contour differences between the real-time shape image and the standard shape image; Generate a difference analysis report based on the contour differences, including the location of the differences, deformation type, deformation degree, and standardized cross-correlation coefficient.

[0012] Preferably, the specific steps for generating the difference analysis report are as follows: The deformation type of the contour difference area is determined by connected component analysis. The deformation type includes flattening, elongation, asymmetry, and burrs. The Euclidean distance method is used to calculate the degree of deformation of the contour difference areas. The location, deformation type, deformation degree of the contour difference areas and the standardized cross-correlation coefficient are integrated to generate a difference analysis report.

[0013] Preferably, the execution steps of the dripping control unit include: The system receives a difference analysis report generated by the shape difference recognition unit and performs an anomaly level assessment, where the anomaly level includes slight, moderate, and severe. The optimal dripping control strategy is matched from the preset control strategy library based on the anomaly level, deformation type, and deformation degree. The optimal dripping control strategy is converted into control commands and sent to the dripping head and / or concentration adjustment mechanism.

[0014] Compared with the prior art, the beneficial effects of the present invention are: This invention discloses an intelligent dripping control system for pharmaceutical production. Before pharmaceutical production begins, a basic information acquisition unit collects relevant data on raw materials, dripping heads, and the environment corresponding to the currently pre-started production line, and integrates them into forming features. These forming features serve as input to a deep learning network, which generates a standard shape image of the droplet. During real-time dripping, a real-time shape image of the droplet is acquired, and the standard shape image is compared with the real-time shape image. A shape difference recognition unit identifies the differences and obtains a difference analysis report. The dripping control unit then adjusts the dripping speed, dripping amount, raw material ratio, and concentration based on the difference analysis and matching of the corresponding control strategy to ensure the quality of pharmaceutical production. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only preferred embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of an intelligent dripping control system for pharmaceutical production according to the present invention.

[0017] In the diagram, 1 is the basic information acquisition unit; 2 is the droplet shape prediction unit; 3 is the real-time shape acquisition unit; 4 is the shape difference recognition unit; 5 is the droplet control unit; 6 is the forming time determination unit; and 7 is the image processing unit. Detailed Implementation

[0018] To better understand the technical content of this invention, a specific embodiment is provided below, and the invention will be further described in conjunction with the accompanying drawings.

[0019] See Figure 1 The present invention provides an intelligent dripping control system for pharmaceutical production, comprising: The basic information acquisition unit 1 is used to collect data on the type of raw material to be added, the concentration of the raw material, the required dripping flow rate, the model of the dripping head, and the dripping environment, and integrate them into the forming features; The dropping shape prediction unit 2 is used to process the forming parameters using a pre-built deep learning network to obtain a standard shape image of the raw material dropping; The real-time shape acquisition unit 3 is used to acquire real-time shape images of the droplets formed by the raw materials added by the dispensing head; The shape difference recognition unit 4 is used to perform difference recognition between standard shape images and real-time shape images, and obtain a difference analysis report between the real-time shape image and the standard shape image; The dripping control unit 5 is used to determine the control strategy based on the difference analysis report, and to control the dripping head and / or concentration adjustment mechanism according to the control strategy; The basic information acquisition unit 1 is connected to the drop shape prediction unit 2, and the shape difference recognition unit 4 is connected to the drop shape prediction unit 2, the real-time shape acquisition unit 3, and the drop control unit 5 respectively.

[0020] This invention discloses an intelligent dripping control system for pharmaceutical production, applied to the dripping control of raw materials in the pharmaceutical production process. It precisely adds raw materials using appropriate dripping heads, ensuring accurate dripping volume and thus guaranteeing that the quality of the produced pharmaceuticals meets expected requirements. However, over long-term use, the dripping heads or other production equipment may experience minor aging or wear, leading to deviations in the dripping process and affecting the quality of the produced pharmaceuticals. To ensure the quality of pharmaceutical processing, this invention compares the real-time dripped droplets with standard droplets and identifies the differences. These differences guide adjustments to the production equipment, achieving monitoring and control of the pharmaceutical production process.

[0021] Before drug production, relevant parameters of the dripping process are collected through the basic information acquisition unit 1, including the type and concentration of raw materials. Different types of raw materials have different physical properties, and raw materials of different concentrations also have different viscosity. In addition to collecting parameters of the raw materials, parameters related to the dripping head are also collected, including the model of the dripping head and the required dripping flow rate. Different flow rates will result in differences in the shape and time of droplet formation. In addition, different models of dripping heads have different orifice diameters, which will also affect the formation of droplets. At the same time, the temperature and humidity of the dripping environment will also cause differences in the shape of the droplets. After collecting the above multi-dimensional data, it can be integrated into the forming feature, which can be input into the dripping shape pre-processing. In measurement unit 2, a pre-built deep learning network is built into the droplet shape prediction unit 2. The deep learning network can process the forming parameters and obtain a standard shape image containing a standard droplet based on the forming parameters. The standard shape image is the normal shape of the droplet that should be generated by the corresponding droplet system under the current set parameters and environment. The generated standard shape image can be compared with the real-time shape image. The real-time shape image is obtained by data acquisition through the real-time shape acquisition unit 3. The real-time shape acquisition unit 3 acquires images of the entire process of droplet generation by the droplet head and selects the most suitable single frame image as the real-time shape image. Finally, the real-time shape image and the standard shape image are input into the shape difference recognition unit 4.

[0022] After receiving the real-time shape image and the standard shape image, the shape difference recognition unit 4 can perform difference recognition, including the deformed parts and the specific deformation amount, and generate a difference analysis report based on the difference recognition results. The difference analysis report can be sent to the drop addition control unit 5. After recognizing the difference analysis report, the drop addition control unit 5 can determine the cause of the difference and determine the control strategy based on the cause of the difference. Finally, the production equipment is adjusted according to the control strategy to ensure that the generated droplets meet the requirements. For example, if it is determined that the drop addition is too much, the drop addition head can be controlled to reduce the drop addition. If the concentration is too high, the concentration adjustment mechanism can be controlled to adjust the ratio of raw materials to improve the concentration, thereby generating real-time droplets that are close to or even identical to the standard droplets.

[0023] Preferably, the execution steps of the basic information acquisition unit 1 include: Obtain the drug production requirements, and extract the types of raw materials, concentrations of raw materials, required dripping flow rates, and models of dripping heads from these requirements. Temperature, humidity, and air pressure data of the environment where the dispensing head is located are collected by the set temperature and humidity sensors and air pressure sensors; Feature extraction was performed on raw material type, raw material concentration, required dripping flow rate, dripping head model, temperature and humidity data, and air pressure data, respectively. The extracted features are concatenated into vector features, and the vector features are output as shaped features.

[0024] Before pharmaceutical production and processing, the process follows corresponding production and processing standards. These standards are dynamically adjusted based on the pharmaceutical production needs, including the type of pharmaceutical being processed, the types of raw materials required, the concentration of the raw materials, and the required dripping flow rate. The production requirements also specify the production equipment used, including the specific model of the dripping head. Additionally, temperature, humidity, and air pressure sensors installed in the production workshop can collect temperature, humidity, and air pressure data of the dripping head's working environment. After extracting the aforementioned multi-dimensional data, preprocessing and feature extraction can be performed. Finally, the extracted features are stitched together to obtain the forming features, which can then be used as input for the subsequent dripping shape prediction unit 2 and forming time determination unit 6.

[0025] Preferably, the execution steps of the droplet shape prediction unit 2 include: Receive and parse the forming features transmitted from the basic information acquisition unit 1; The parsed shaped features are then input into a pre-constructed convolutional neural network; The shaping parameters are subjected to multi-layer nonlinear transformation and feature extraction through the forward propagation process of a convolutional neural network; The output layer of the convolutional neural network outputs standard image data containing the standard shape of a droplet.

[0026] The convolutional neural network is trained using a large amount of historical data, including droplet shape images and multi-dimensional historical data such as raw materials, dropper heads, and environment during droplet generation. After matching the droplet shape images and multi-dimensional historical data, the data is divided into training, validation, and test sets in a 7:2:1 ratio. The convolutional neural network can then be trained and put into use. Its input is the forming features integrated by the basic information acquisition unit 1. After parsing the forming features, multi-layer nonlinear transformations and feature extraction are performed through its forward propagation process. Then, the network output layer generates a two-dimensional pixel matrix representing the ideal droplet shape, i.e., standard image data. The standard image data can be used for subsequent shape difference recognition.

[0027] Preferably, it also includes a forming time determination unit 6, which is used to determine the forming time of the droplet according to the forming characteristics and to transmit the forming time to the real-time shape acquisition unit 3. The forming time determination unit 6 is connected to the basic information acquisition unit 1 and the real-time shape acquisition unit 3 respectively.

[0028] The formation of a droplet is a continuous process, but there is a point where the droplet is fully formed. The droplet in the standard shape image is the droplet at the point of full formation. The image data acquired by the real-time shape acquisition unit 3 is an image sequence. It is necessary to extract the real-time shape of the fully formed point from the image sequence. The forming time determination unit 6 is used to determine the time when the droplet is fully formed. It can predict the forming time based on the raw material concentration, type, viscosity, flow rate and environmental parameters in the forming features. After the forming time is transmitted to the real-time shape acquisition unit 3, the real-time shape acquisition unit 3 can extract the real-time shape image based on the forming time.

[0029] Preferably, the execution steps of the real-time shape acquisition unit 3 include: A high-speed camera is deployed near the dripping head to collect image data of the entire process of raw material dripping from the dripping head. The entire process image data is decomposed frame by frame to obtain the starting frame image; Based on the forming time, the forming frame image is obtained by shifting backward from the starting frame image, and the forming frame image is output as a real-time shape image of the droplet.

[0030] A high-speed camera deployed near the droplet head can collect image data during the dropleting process, including the entire process from the initial appearance of the droplet to its complete formation and falling. Then, the entire process image data is decomposed frame by frame. First, the moment when the droplet begins to appear is determined. Then, based on the formation time determination unit 6, the formation time is shifted backward to obtain the formation frame image. The droplet in the formation frame image is the droplet at the moment of complete formation. It can be used to compare with the standard droplet in the standard shape image. Finally, the formation frame image is sent as a real-time shape image to the shape difference recognition unit 4.

[0031] Preferably, it also includes an image processing unit 7, which is used to perform noise reduction, grayscale conversion and edge enhancement processing on the standard shape image and the real-time shape image. The image processing unit 7 is connected to the drop-on shape prediction unit 2, the real-time shape acquisition unit 3 and the shape difference recognition unit 4 respectively.

[0032] After obtaining the real-time shape image and the standard shape image, in order for the subsequent shape difference recognition unit 4 to accurately identify the shape, it is necessary to perform image processing on the real-time shape image and the standard shape image, including noise reduction, grayscale conversion and edge enhancement, etc., in order to obtain a clearly visible droplet shape image.

[0033] Preferably, the execution steps of the shape difference recognition unit 4 include: The SIFT algorithm was used to extract the contour feature points of the droplets in real-time shape images and standard shape images; Two sets of contour feature points are initially matched by a fast approximate nearest neighbor search, and mismatched points after the initial matching are removed to obtain accurately matched feature pairs. The precisely matched image is input into a lightweight CNN model, and the standardized cross-correlation coefficient is calculated on the feature map output by the lightweight CNN model. The standardized cross-correlation coefficient is compared with a preset threshold to identify and mark the contour differences between the real-time shape image and the standard shape image; Generate a difference analysis report based on the contour differences, including the location of the differences, deformation type, deformation degree, and standardized cross-correlation coefficient.

[0034] The shape difference recognition unit 4 incorporates the SIFT algorithm and a lightweight CNN model. It uses the SIFT algorithm to extract features from the droplets in both real-time and standard shape images, obtaining several sets of contour feature points. These contour feature points contain coordinate, scale, and orientation information. Then, a FLANN (Fast Approximate Nearest Neighbor Search) matcher is used to perform preliminary matching on the two sets of contour feature points. For the preliminarily matched contour feature points, the RANSAC algorithm is used to remove mismatches, resulting in precisely matched feature pairs. The precisely matched real-time and standard shape images can then be input into the lightweight CNN model. The feature fusion layer enhances the expression of differential features. The lightweight CNN model can generate feature maps, and then calculate the standardized cross-correlation coefficient (NCC) on the feature maps. The NCC can quantify the similarity of the contours. By comparing the NCC with a preset threshold, the contour difference between the real-time shape image and the standard shape image can be identified. Then, a difference analysis report can be generated based on the contour difference. The difference analysis report includes the specific location of the difference, the type of droplet deformation, the degree of deformation, and the corresponding NCC value. The difference analysis report can be sent to the droplet control unit 5, which will parse and control the production equipment.

[0035] Preferably, the specific steps for generating the difference analysis report are as follows: The deformation type of the contour difference area is determined by connected component analysis. The deformation type includes flattening, elongation, asymmetry, and burrs. The Euclidean distance method is used to calculate the degree of deformation of the contour difference areas. The location, deformation type, deformation degree of the contour difference areas and the standardized cross-correlation coefficient are integrated to generate a difference analysis report.

[0036] The location of the difference can be determined by the specific pixels of the contour difference area. Then, the deformation type and degree of the contour difference area need to be evaluated. The deformation types include flattening, elongation, asymmetry, and burrs. Flattening means that there may be problems with low surface tension of raw materials or slow dripping flow rate. Elongation means that there may be problems with high viscosity / concentration of raw materials or excessive dripping flow rate. Asymmetry indicates that there may be abnormalities in the dripping head, such as local blockage of the dripping head, wear and deformation of the inner wall, or crystals / residual adhesion on the surface of the dripping head. Burrs are mostly due to uneven mixing of raw materials, or trace crystal precipitation of the dripping head, or contamination of the inner wall. Different deformation types correspond to different problems in the production equipment. The control strategy generated by the subsequent dripping control unit 5 will also be matched accordingly based on the different deformation types.

[0037] In addition, the degree of deformation of the contour difference area will be calculated by Euclidean distance. The degree of deformation is quantified in pixels. The greater the degree of deformation, the more serious the problem. Finally, the above contents are integrated into a difference analysis report and sent to the drop addition control unit 5.

[0038] Preferably, the execution steps of the dripping control unit 5 include: The anomaly level assessment is performed on the difference analysis report generated by the shape difference recognition unit 4, which includes slight, moderate, and severe anomalies. The optimal dripping control strategy is matched from the preset control strategy library based on the anomaly level, deformation type, and deformation degree. The optimal dripping control strategy is converted into control commands and sent to the dripping head and / or concentration adjustment mechanism.

[0039] The dispensing control unit 5 analyzes the difference analysis report and assesses the anomaly level based on the deformation type and degree. The higher the anomaly level, the more serious the problem. The generated control strategy requires not only automation but also manual monitoring and closed-loop feedback. After determining the anomaly level, the optimal dispensing control strategy is matched according to the anomaly level, deformation type, and degree. The optimal dispensing control strategy is then converted into control commands to control the dispensing head and / or concentration adjustment mechanism. For example, when the strategy involves the dispensing head, control commands containing specific adjustment parameters can be generated and sent to the dispensing head's actuator motor or piezoelectric ceramic driver. When the strategy involves concentration adjustment, control commands are generated and sent to the control valve or micro-pump of the concentration adjustment mechanism to change the solvent or raw material ratio. After issuing the control command, a feedback monitoring cycle is triggered to wait for and evaluate the changes in the subsequent real-time shape image until the difference is eliminated.

[0040] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An intelligent dripping control system for pharmaceutical production, characterized in that, include: The basic information acquisition unit is used to collect data on the type of raw material to be added, raw material concentration, required dripping flow rate, dripping head model, and dripping environment, and integrate them into forming features; The droplet shape prediction unit is used to process the forming parameters using a pre-built deep learning network to obtain a standard shape image of the raw material droplet; The real-time shape acquisition unit is used to acquire real-time shape images of the droplets formed by the raw materials being added by the dispensing head; The shape difference recognition unit is used to perform difference recognition between standard shape images and real-time shape images, and obtain a difference analysis report between the real-time shape image and the standard shape image; The dispensing control unit is used to determine the control strategy based on the difference analysis report and to control the dispensing head and / or concentration adjustment mechanism according to the control strategy; The basic information acquisition unit is connected to the drop shape prediction unit, and the shape difference recognition unit is connected to the drop shape prediction unit, the real-time shape acquisition unit, and the drop control unit, respectively.

2. The intelligent dripping control system for pharmaceutical production according to claim 1, characterized in that, The execution steps of the basic information collection unit include: Obtain the drug production requirements, and extract the types of raw materials, concentrations of raw materials, required dripping flow rates, and models of dripping heads from these requirements. Temperature, humidity, and air pressure data of the environment where the dispensing head is located are collected by the set temperature and humidity sensors and air pressure sensors; Feature extraction was performed on raw material type, raw material concentration, required dripping flow rate, dripping head model, temperature and humidity data, and air pressure data, respectively. The extracted features are concatenated into vector features, and the vector features are output as shaped features.

3. The intelligent dripping control system for pharmaceutical production according to claim 1, characterized in that, The execution steps of the droplet shape prediction unit include: Receive and parse the formed features transmitted from the basic information acquisition unit; The parsed shaped features are then input into a pre-constructed convolutional neural network; The shaping parameters are subjected to multi-layer nonlinear transformation and feature extraction through the forward propagation process of a convolutional neural network; The output layer of the convolutional neural network outputs standard image data containing the standard shape of a droplet.

4. The intelligent dripping control system for pharmaceutical production according to claim 1, characterized in that, It also includes a forming time determination unit, which is used to determine the forming time of the droplet based on the forming characteristics and input the forming time to the real-time shape acquisition unit. The forming time determination unit is connected to the basic information acquisition unit and the real-time shape acquisition unit respectively.

5. The intelligent dripping control system for pharmaceutical production according to claim 4, characterized in that, The execution steps of the real-time shape acquisition unit include: A high-speed camera is deployed near the dripping head to collect image data of the entire process of raw material dripping from the dripping head. The entire process image data is decomposed frame by frame to obtain the starting frame image; Based on the forming time, the forming frame image is obtained by shifting backward from the starting frame image, and the forming frame image is output as a real-time shape image of the droplet.

6. The intelligent dripping control system for pharmaceutical production according to claim 1, characterized in that, It also includes an image processing unit for denoising, grayscale conversion, and edge enhancement of standard shape images and real-time shape images. The image processing unit is connected to the droplet shape prediction unit, the real-time shape acquisition unit, and the shape difference recognition unit.

7. The intelligent dripping control system for pharmaceutical production according to claim 1, characterized in that, The execution steps of the shape difference recognition unit include: The SIFT algorithm was used to extract the contour feature points of the droplets in real-time shape images and standard shape images; Two sets of contour feature points are initially matched by a fast approximate nearest neighbor search, and mismatched points after the initial matching are removed to obtain accurately matched feature pairs. The precisely matched image is input into a lightweight CNN model, and the standardized cross-correlation coefficient is calculated on the feature map output by the lightweight CNN model. The standardized cross-correlation coefficient is compared with a preset threshold to identify and mark the contour differences between the real-time shape image and the standard shape image; Generate a difference analysis report based on the contour differences, including the location of the differences, deformation type, deformation degree, and standardized cross-correlation coefficient.

8. The intelligent dripping control system for pharmaceutical production according to claim 7, characterized in that, The specific steps for generating the difference analysis report are as follows: The deformation type of the contour difference area is determined by connected component analysis. The deformation type includes flattening, elongation, asymmetry, and burrs. The Euclidean distance method is used to calculate the degree of deformation of the contour difference areas. The location, deformation type, deformation degree of the contour difference areas and the standardized cross-correlation coefficient are integrated to generate a difference analysis report.

9. The intelligent dripping control system for pharmaceutical production according to claim 7, characterized in that, The execution steps of the dripping control unit include: The system receives a difference analysis report generated by the shape difference recognition unit and performs an anomaly level assessment, where the anomaly level includes slight, moderate, and severe. The optimal dripping control strategy is matched from the preset control strategy library based on the anomaly level, deformation type, and deformation degree. The optimal dripping control strategy is converted into control commands and sent to the dripping head and / or concentration adjustment mechanism.

Images