Intelligent control method and system of equipment applied to tablet defect detection

Abstract

The application relates to the technical field of artificial intelligence, and discloses a device intelligent control method and system applied to tablet defect detection, which comprises the following steps: extracting data features of tablet data to obtain tablet data features; analyzing high-order feature vectors of the tablet data features, analyzing defect parameters of a tablet to be detected, marking a preliminary defect tablet in the tablet to be detected corresponding to the defect parameters, collecting multi-angle images of the preliminary defect tablet, performing pixel-level segmentation on the multi-angle images to obtain segmented multi-angle images; calculating a disintegration time limit and a dynamic friability index of the preliminary defect tablet to determine the defect severity of the preliminary defect tablet; constructing a defect cause analysis atlas of the preliminary defect tablet to generate device defect optimization parameters of a tablet production scene, and performing device intelligent control of the tablet production scene based on the device defect optimization parameters. The application can improve the accuracy of tablet defect detection.

Description

Technical Field

[0001] This invention relates to an intelligent control method and system for tablet defect detection, belonging to the field of artificial intelligence technology. Background Technology

[0002] Tablet defect detection refers to the quality inspection of manufactured tablets (pills) in the pharmaceutical industry to identify and reject products that do not meet quality standards or are defective. Through these detection methods, pharmaceutical companies can ensure that every tablet meets strict quality standards, thereby protecting the safety of patients' medication.

[0003] Traditional methods for detecting tablet defects mainly rely on manual visual inspection or machine vision systems based on a single modality (such as two-dimensional images). Manual inspection is inefficient, prone to fatigue, and highly subjective, making it difficult to guarantee consistency. While machine vision based on two-dimensional images improves efficiency, it can only capture surface information and cannot perceive the internal structure, three-dimensional morphology, or dynamic changes during the production process (such as sound characteristics) of the tablet, resulting in low accuracy in tablet defect detection. Summary of the Invention

[0004] This invention provides an intelligent control method and system for tablet defect detection, the main purpose of which is to improve the accuracy of tablet defect detection.

[0005] To achieve the above objectives, the present invention provides an intelligent control method for equipment applied to tablet defect detection, comprising:

[0006] Tablet data of tablets to be tested in a tablet production scenario is collected, wherein the tablet data includes surface images, three-dimensional morphology data and acoustic signals of the tableting process, and the data features of the tablet data are extracted to obtain tablet data features;

[0007] The high-order feature vector of the tablet data features is analyzed using a multimodal feature interaction model in a pre-constructed defect screening network. Based on the high-order feature vector, the defect parameters of the tablet to be tested are analyzed using a defect classification model in the defect screening network. The defect parameters include a preliminary defect type label and a preliminary defect location.

[0008] The defect parameters are labeled to correspond to the preliminary defective tablets in the tablets to be tested, and multi-angle images of the preliminary defective tablets are acquired. Based on the defect parameters, the multi-angle images are segmented at the pixel level to obtain segmented multi-angle images.

[0009] Based on the segmented multi-angle images, the disintegration time and dynamic fragility index of the initially defective tablet are calculated to determine the severity of the defect in the initially defective tablet.

[0010] Based on the severity of the defect, a defect cause analysis map of the preliminary defective tablet is constructed to generate equipment defect optimization parameters for the tablet production scenario. Based on the equipment defect optimization parameters, intelligent equipment control for the tablet production scenario is executed.

[0011] Optionally, the step of extracting data features from the tablet data to obtain tablet data features includes:

[0012] Calculate the extrinsic matrix of the surface image and three-dimensional morphology data in the tablet data;

[0013] Based on the extrinsic parameter matrix, a mapping three-dimensional point cloud coordinate system is established between the surface image and the three-dimensional topography data;

[0014] Establish a sound source-image region mapping table for the acoustic signals of the tableting process in the surface image and the tablet data;

[0015] Based on the mapped 3D point cloud coordinate system and the sound source-image region mapping table, the tablet data is spatially registered to obtain registered tablet data, wherein the registered tablet data includes: registered surface image, registered 3D morphology data and registered acoustic signal of tableting process;

[0016] Multi-scale feature maps, point cloud global features, and time-frequency domain joint features of the registration surface image, registration three-dimensional topography data, and acoustic signals of the registration pressing process are extracted respectively.

[0017] Based on the multi-scale feature map, point cloud global features, and time-frequency domain joint features, the tablet data features of the tablet data are determined.

[0018] Optionally, the step of extracting the multi-scale feature maps, point cloud global features, and time-domain-frequency domain joint features of the registration surface image, registration three-dimensional topography data, and acoustic signals from the registration pressing process includes:

[0019] The registered surface image is downsampled to obtain a feature map sequence;

[0020] The Scharr gradient operator of the feature map sequence is calculated to extract multi-scale feature maps of the sampled and registered surface image;

[0021] The registered 3D topography data is divided into a voxel mesh;

[0022] Identify the principal curvature direction of the tablet in the voxel mesh to perform voxel reconstruction on the voxel mesh, thereby obtaining a reconstructed voxel mesh;

[0023] The reconstructed voxel mesh is subjected to density-curvature joint filtering to obtain a filtered reconstructed voxel mesh;

[0024] Calculate the seven-dimensional feature vector of the filtered and reconstructed voxel mesh to determine the global point cloud features of the filtered and reconstructed voxel mesh;

[0025] Analyze the Mel spectrum of the acoustic signal during the registration and tableting process;

[0026] Based on the Mel spectrum, identify the time-frequency joint characteristics of the acoustic signal during the registration and tableting process.

[0027] Optionally, the analysis of the Mel spectrum of the acoustic signal during the registration and tableting process includes:

[0028] Analyze the minimum and maximum frequencies of the acoustic signals during the registration and tableting process;

[0029] Analyze the resonant peak shift coefficient and nonlinear intensity coefficient of the acoustic signal during the registration and compression process;

[0030] Based on the minimum frequency of the frequency band, the maximum frequency of the frequency band, the resonant peak shift coefficient, and the nonlinear intensity coefficient, the filter center frequency of the acoustic signal during the registration and pressing process is calculated using the following formula:

[0031] ;

[0032] in, The acoustic signal representing the registration and tableting process. The center frequency of each filter. Indicates the minimum frequency of the frequency band. Indicates the maximum frequency of the frequency band. The acoustic signal representing the registration and tableting process. One filter, This indicates the number of filters used to transmit the acoustic signals during the registration and tableting process. This represents the resonance peak shift coefficient. Represents the nonlinear strength coefficient. Represents the arctangent function;

[0033] Based on the center frequency of the filter, the Mel spectrum of the acoustic signal during the registration and tableting process was analyzed.

[0034] Optionally, the step of analyzing the high-order feature vectors of the tablet data features using a pre-constructed multimodal feature interaction model in the defect screening network includes:

[0035] Calculate the gating weights of the time-frequency joint features in the tablet data features;

[0036] Based on the gating weights, the attention weights of the multi-scale feature map and the global point cloud features in the tablet data features are calculated using the attention layer in the multimodal feature interaction model.

[0037] Based on the attention weights, the tablet data features are aggregated using the aggregation layer in the multimodal feature interaction model to obtain aggregated features;

[0038] The aggregated features are reduced in dimensionality using the feature dimensionality reduction layer in the multimodal feature interaction model to obtain the high-order feature vector of the tablet data features.

[0039] Optionally, the step of performing pixel-level segmentation on the multi-angle image based on the defect parameters to obtain a segmented multi-angle image includes:

[0040] Based on the defect parameters, weakly supervised labels are generated for the multi-angle images to construct an interpretable segmentation network for the multi-angle images.

[0041] Calculate the viewpoint homography matrix of the multi-angle image;

[0042] Based on the interview homography matrix, the multi-angle images are spatially aligned to obtain aligned multi-angle images;

[0043] The illustrative segmentation network is used to perform pixel-level segmentation on the aligned multi-angle image to obtain the segmented multi-angle image.

[0044] Optionally, calculating the disintegration time and dynamic fragility index of the initially defective tablet based on the segmented multi-angle images includes:

[0045] The defect volume of the preliminary defective tablet is calculated based on the segmented multi-angle image.

[0046] Based on the defect volume, the permeation path of the initially defective tablet is analyzed;

[0047] Based on the described penetration path, the initially defective tablets are classified into open defects and closed defects;

[0048] Analyze the defect cross-sectional area and total cross-sectional area of ​​the initially defective tablet;

[0049] Based on the defect cross-sectional area and the total cross-sectional area of ​​the tablet, the open-type defect disintegration time of the initially defective tablet is calculated;

[0050] Calculate the closed-type defect disintegration time of the initially defective tablet;

[0051] Based on the disintegration time limit of the open-type defect and the disintegration time limit of the closed-type defect, the disintegration time limit of the initially defective tablet is determined;

[0052] Based on the segmented multi-angle images, the radius of curvature of the defect tip of the initially defective tablet is analyzed to calculate the maximum stress of the initially defective tablet;

[0053] Calculate the dynamic brittleness index of the initially defective tablet based on the maximum stress.

[0054] Optionally, calculating the open-type defect disintegration time of the initially defective tablet based on the defect cross-sectional area and the total cross-sectional area of ​​the tablet includes:

[0055] Identify the liquid viscosity and concentration gradient of the dissolution medium corresponding to the initially defective tablet;

[0056] Analyze the drug diffusion coefficient of the initially defective tablet in the dissolution medium;

[0057] Based on the liquid viscosity, the concentration gradient, the drug diffusion coefficient, the defect cross-sectional area, and the total cross-sectional area of ​​the tablet, the disintegration time of the open-type defect corresponding to the initially defective tablet is calculated.

[0058] Optionally, constructing the defect cause analysis map of the preliminary defective tablet based on the defect severity includes:

[0059] Obtain the defect-related process data of the preliminary defective tablet;

[0060] Based on the defect-related process data, a defect-process multidimensional feature vector is constructed for the preliminary defect tablet.

[0061] Calculate the causal correlation weights of the defect-process multidimensional feature vector;

[0062] Based on the causal correlation weights and the defect severity, a defect causal analysis map of the preliminary defective tablet is constructed.

[0063] To address the aforementioned problems, the present invention also provides an intelligent control system for tablet defect detection, the system comprising:

[0064] The tablet data feature extraction module is used to collect tablet data of tablets to be tested in a tablet production scenario. The tablet data includes surface images, three-dimensional morphology data and acoustic signals of the tableting process. The module extracts the data features of the tablet data to obtain tablet data features.

[0065] The preliminary defect analysis module is used to analyze the high-order feature vector of the tablet data features using the multimodal feature interaction model in the pre-constructed defect screening network, and based on the high-order feature vector, to analyze the defect parameters of the tablet to be tested using the defect classification model in the defect screening network, wherein the defect parameters include preliminary defect type label and preliminary defect location.

[0066] A multi-angle image segmentation module is used to mark the defect parameters corresponding to the preliminary defect tablets in the tablets to be detected, and to acquire multi-angle images of the preliminary defect tablets. Based on the defect parameters, the multi-angle images are segmented at the pixel level to obtain segmented multi-angle images.

[0067] The defect severity analysis module is used to calculate the disintegration time and dynamic fragility index of the initially defective tablet based on the segmented multi-angle images, so as to determine the defect severity of the initially defective tablet;

[0068] The equipment intelligent control module is used to construct a defect cause analysis map of the preliminary defective tablet based on the defect severity, generate equipment defect optimization parameters for the tablet production scenario, and execute equipment intelligent control for the tablet production scenario based on the equipment defect optimization parameters.

[0069] First, by fusing multimodal data such as surface images, 3D morphology, and tablet compression acoustic signals, this system can acquire more comprehensive and in-depth product information than traditional single-modal detection. This greatly improves the accuracy and comprehensiveness of defect detection, enabling the identification of complex defects that were previously difficult to detect, such as internal microcracks and density inhomogeneities, effectively reducing the false negative rate and ensuring the quality and safety of pharmaceuticals. Second, the solution goes beyond defect identification; it extracts high-order features through a multimodal feature interaction model and uses a defect classification model for preliminary analysis, obtaining preliminary parameters including defect type and location. This lays the foundation for subsequent detailed analysis. Subsequently, through pixel-level segmentation of multi-angle images and calculation of key quality attributes (such as disintegration time and friability), The system can quantitatively assess the severity of defects, achieving a qualitative leap from "whether there is a defect" to "how severe the defect is," providing a more scientific basis for subsequent quality grading or handling. More innovative and valuable is its close integration of defect analysis with production process optimization. The causal analysis map built based on defect severity can trace the possible root causes of defects, whether they are raw material problems, formulation issues, or improper equipment parameter settings. The resulting optimized equipment defect parameters are then used to execute intelligent equipment control, forming a closed-loop system from detection to feedback and optimization. This not only improves the pass rate of single batches of products but also enhances the stability and consistency of the entire production process by continuously optimizing production equipment parameters, reducing the scrap rate. Therefore, this invention can improve the accuracy of tablet defect detection. Attached Figure Description

[0070] Figure 1 This is a flowchart illustrating an intelligent control method for tablet defect detection provided in an embodiment of the present invention.

[0071] Figure 2 This is a schematic diagram of a module for implementing the intelligent control method for tablet defect detection according to an embodiment of the present invention.

[0072] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0073] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0074] This application provides an intelligent control method for tablet defect detection. The executing entity of this intelligent control method for tablet defect detection includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the intelligent control method for tablet defect detection can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster.

[0075] Example 1:

[0076] Reference Figure 1 The diagram shown is a flowchart illustrating an intelligent control method for tablet defect detection provided in an embodiment of the present invention. In this embodiment, the intelligent control method for tablet defect detection includes:

[0077] S1. Collect tablet data of the tablets to be tested in the tablet production scenario, wherein the tablet data includes surface images, three-dimensional morphology data and acoustic signals of the tableting process, and extract the data features of the tablet data to obtain tablet data features.

[0078] It should be explained that the tablet production scenario refers to the physical environment in which tablets are produced through a series of production processes, from raw material mixing, granulation, and drying to final tableting, coating, and packaging. The tablet to be tested refers to an individual tablet that has just been pressed by the tableting machine on the production line and has not yet undergone final quality inspection or packaging, and needs to be systematically checked for defects. The surface image refers to the two-dimensional (2D) visual information obtained by taking pictures of the outer surface of the tablet to be tested using a camera (usually a high-resolution industrial camera). The three-dimensional morphology data refers to the measurement of the surface of the tablet to be tested using specific sensors or scanning technologies (such as structured light scanning, laser scanning, or stereoscopic vision-based methods) to obtain its three-dimensional spatial coordinate information. The acoustic signal of the tableting process refers to the sound information collected by a microphone or other acoustic sensors during the dynamic process of the tablet being pressed by the tableting machine.

[0079] The present invention extracts data features from the tablet data, and the resulting tablet data features can obtain more comprehensive and in-depth product information than traditional single-modality detection.

[0080] Specifically, the extraction of data features from the tablet data to obtain tablet data features includes:

[0081] Calculate the extrinsic matrix of the surface image and three-dimensional morphology data in the tablet data;

[0082] Based on the extrinsic parameter matrix, a mapping three-dimensional point cloud coordinate system is established between the surface image and the three-dimensional topography data;

[0083] Establish a sound source-image region mapping table for the acoustic signals of the tableting process in the surface image and the tablet data;

[0084] Based on the mapped 3D point cloud coordinate system and the sound source-image region mapping table, the tablet data is spatially registered to obtain registered tablet data, wherein the registered tablet data includes: registered surface image, registered 3D morphology data and registered acoustic signal of tableting process;

[0085] Multi-scale feature maps, point cloud global features, and time-frequency domain joint features of the registration surface image, registration three-dimensional topography data, and acoustic signals of the registration pressing process are extracted respectively.

[0086] Based on the multi-scale feature map, point cloud global features, and time-frequency domain joint features, the tablet data features of the tablet data are determined.

[0087] Wherein, the extrinsic parameter matrix refers to the relative position and orientation between the camera or sensor (the camera that captures the surface image and the scanner that acquires the three-dimensional topography data), the mapped three-dimensional point cloud coordinate system refers to a unified three-dimensional coordinate system, by applying the extrinsic parameter matrix, the two-dimensional pixel coordinates in the surface image and the three-dimensional point coordinates in the three-dimensional topography data are transformed to this unified coordinate system, the sound source-image region mapping table refers to the data structure that records which region on the surface image corresponds to the acoustic features of a specific time period or a specific frequency when the acoustic signals of the tableting process are collected, the registered tablet data refers to the tablet data after spatial coordinate registration of the tablet data, including the surface image, the three-dimensional topography data and the acoustic signals of the tableting process, and the multi-scale feature map refers to the feature representation extracted at different resolutions (scales) of the image. Low-level feature maps capture details such as edges and textures, while high-level feature maps capture more abstract shapes of parts or the whole of an object. The global point cloud features refer to features extracted from 3D point cloud data that can describe the overall shape and distribution of the entire tablet or most of its area. The joint time-domain and frequency-domain features refer to features obtained by analyzing acoustic signals. For example, the mean and variance of the signal (time domain), as well as the main frequencies and spectral energy distribution (frequency domain) can be extracted. The tablet data features include multi-scale feature maps, global point cloud features, and joint time-domain and frequency-domain features.

[0088] Optionally, the extrinsic matrix for calculating the surface image and three-dimensional morphology data in the tablet data can be calculated using common feature points (such as checkerboard corner points) on a calibration board.

[0089] Further, the extraction of multi-scale feature maps, point cloud global features, and time-domain-frequency domain joint features of the registration surface image, registration three-dimensional topography data, and registration pressing process acoustic signals includes:

[0090] The registered surface image is downsampled to obtain a feature map sequence;

[0091] The Scharr gradient operator of the feature map sequence is calculated to extract multi-scale feature maps of the sampled and registered surface image;

[0092] The registered 3D topography data is divided into a voxel mesh;

[0093] Identify the principal curvature direction of the tablet in the voxel mesh to perform voxel reconstruction on the voxel mesh, thereby obtaining a reconstructed voxel mesh;

[0094] The reconstructed voxel mesh is subjected to density-curvature joint filtering to obtain a filtered reconstructed voxel mesh;

[0095] Calculate the seven-dimensional feature vector of the filtered and reconstructed voxel mesh to determine the global point cloud features of the filtered and reconstructed voxel mesh;

[0096] Analyze the Mel spectrum of the acoustic signal during the registration and tableting process;

[0097] Based on the Mel spectrum, identify the time-frequency joint characteristics of the acoustic signal during the registration and tableting process.

[0098] The feature map sequence refers to a set of images with decreasing resolution generated by multi-level downsampling (such as Gaussian pyramids) of the registered surface image; the Scharr gradient operator refers to an improved 3×3 edge detection convolution kernel with better rotational symmetry than the traditional Sobel operator; the voxel mesh refers to a discretized representation that divides the 3D point cloud data into regular cubic units (voxels); the principal curvature direction of the tablet refers to the maximum and minimum curvature directions of the local surface of the point cloud calculated by PCA (principal component analysis); and the reconstructed volume... A voxel mesh refers to a new mesh after adjusting the voxel size according to the principal curvature direction. The filtered reconstructed voxel mesh refers to the result after applying density-curvature joint constraints to the reconstructed voxels, retaining voxels that meet the following conditions. The seven-dimensional feature vector refers to the seven geometric features extracted from each filtered voxel, including density, curvature, normal vector angle, height difference, surface roughness, anisotropy index, and connected region size. The Mel spectrum refers to the time-frequency representation generated by the acoustic signal after undergoing short-time Fourier transform (STFT) and passing through a Mel-scale filter bank (simulating the characteristics of human hearing).

[0099] Furthermore, the analysis of the Mel spectrum of the acoustic signal during the registration and tableting process includes:

[0100] Analyze the minimum and maximum frequencies of the acoustic signals during the registration and tableting process;

[0101] Analyze the resonant peak shift coefficient and nonlinear intensity coefficient of the acoustic signal during the registration and compression process;

[0102] Based on the minimum frequency of the frequency band, the maximum frequency of the frequency band, the resonant peak shift coefficient, and the nonlinear intensity coefficient, the filter center frequency of the acoustic signal during the registration and pressing process is calculated using the following formula:

[0103] ;

[0104] in, The acoustic signal representing the registration and tableting process. The center frequency of each filter. Indicates the minimum frequency of the frequency band. Indicates the maximum frequency of the frequency band. The acoustic signal representing the registration and tableting process. One filter, This indicates the number of filters used to transmit the acoustic signals during the registration and tableting process. This represents the resonance peak shift coefficient. Represents the nonlinear strength coefficient. Represents the arctangent function;

[0105] Based on the center frequency of the filter, the Mel spectrum of the acoustic signal during the registration and compression process is analyzed.

[0106] Wherein, the minimum frequency of the frequency band refers to the lower limit of the frequency range of interest when analyzing the acoustic signal of the registration and pressing process; the maximum frequency of the frequency band refers to the upper limit of the frequency range of interest when analyzing the acoustic signal of the registration and pressing process; the resonant peak offset coefficient is used to adjust the center frequency of the filter to adapt to or compensate for the systematic offset of the resonant peak frequency that may exist in the acoustic signal, and in this invention it can be 0.45; the nonlinear intensity coefficient refers to the intensity of the nonlinear function used to control the calculation of the center frequency of the filter, and in this invention it can be 3.0.

[0107] S2. Analyze the high-order feature vector of the tablet data features using the multimodal feature interaction model in the pre-constructed defect screening network, and based on the high-order feature vector, analyze the defect parameters of the tablet to be tested using the defect classification model in the defect screening network, wherein the defect parameters include preliminary defect type label and preliminary defect location.

[0108] This invention utilizes a multimodal feature interaction model in a pre-constructed defect screening network to analyze the high-order feature vectors of the tablet data features, providing a basis for subsequent defect analysis.

[0109] In detail, the analysis of the high-order feature vectors of the tablet data features using the multimodal feature interaction model in the pre-constructed defect screening network includes:

[0110] Calculate the gating weights of the time-frequency joint features in the tablet data features;

[0111] Based on the gating weights, the attention weights of the multi-scale feature map and the global point cloud features in the tablet data features are calculated using the attention layer in the multimodal feature interaction model.

[0112] Based on the attention weights, the tablet data features are aggregated using the aggregation layer in the multimodal feature interaction model to obtain aggregated features;

[0113] The aggregated features are reduced in dimensionality using the feature dimensionality reduction layer in the multimodal feature interaction model to obtain the high-order feature vector of the tablet data features.

[0114] The gating weights refer to the degree of influence of joint time-domain and frequency-domain features on subsequent calculations. The multimodal feature interaction model refers to a model that analyzes the complex and nonlinear interaction relationships between features of multiple modalities (such as images, 3D shapes, and acoustic signals). The attention layer refers to a layer that dynamically allocates different "attention" based on the importance or relevance of the input data itself. The attention weights refer to the numerical vectors calculated and output by the attention layer after being normalized by Softmax. The aggregation layer refers to a layer that merges features from different modalities according to the calculated attention weights. The aggregated features refer to the intermediate representation of multimodal features after attention weighting and fusion. The feature dimensionality reduction layer refers to a layer used to reduce the dimensionality of the aggregated features. The higher-order feature vector refers to the final feature representation output by the feature dimensionality reduction layer.

[0115] Optionally, the step of using the feature dimensionality reduction layer in the multimodal feature interaction model to reduce the dimensionality of the aggregated features and obtain the high-order feature vector of the tablet data features is achieved by compressing the 512-dimensional aggregated features into a 128-dimensional high-order vector.

[0116] This invention, based on the aforementioned high-order feature vectors, utilizes a defect classification model within the initial defect screening network to analyze the defect parameters of the tablet under test, laying the foundation for subsequent refined analysis. The defect classification model is the core component responsible for transforming high-order feature vectors into specific defect information (type and location). It is a machine learning model trained using historical high-order defect feature vectors. The preliminary defect type label refers to the initial judgment result output by the defect classification model regarding the type of defect in the tablet under test; for example, if the preset categories include "no defects," "cracked tablet," "loose tablet," and "stuck tablet," the preliminary defect location refers to the preliminary estimate output by the defect classification model regarding the approximate area where the defect appears on the tablet.

[0117] S3. Mark the defect parameters corresponding to the preliminary defect tablets in the tablets to be tested, and acquire multi-angle images of the preliminary defect tablets. Based on the defect parameters, perform pixel-level segmentation on the multi-angle images to obtain segmented multi-angle images.

[0118] It should be explained that the preliminary defective tablets refer to tablet samples marked as having potential defects in the analysis results of the defect screening network, and the multi-angle images refer to a collection of images of the preliminary defective tablets taken from multiple different angles.

[0119] Based on the aforementioned defect parameters, this invention performs pixel-level segmentation on the multi-angle image, obtaining a segmented multi-angle image that provides a foundation for subsequent defect quantification.

[0120] Specifically, the step of performing pixel-level segmentation on the multi-angle image based on the defect parameters to obtain a segmented multi-angle image includes:

[0121] Based on the defect parameters, weakly supervised labels are generated for the multi-angle images to construct an interpretable segmentation network for the multi-angle images.

[0122] Calculate the viewpoint homography matrix of the multi-angle image;

[0123] Based on the interview homography matrix, the multi-angle images are spatially aligned to obtain aligned multi-angle images;

[0124] The illustrative segmentation network is used to perform pixel-level segmentation on the aligned multi-angle image to obtain the segmented multi-angle image.

[0125] The weakly supervised labels are generated based on defect parameters and are used to provide approximate defect region indications for multi-angle images. The interpretable segmentation network is a deep learning network used for pixel-level defect segmentation of multi-angle images. The viewpoint homography matrix is ​​a mathematical matrix used to describe the spatial transformation relationship between images from different viewpoints. The aligned multi-angle images are a set of images that are spatially aligned after the multi-angle images have been transformed by the viewpoint homography matrix. The segmented multi-angle images are the result of the aligned multi-angle images being processed by the interpretable segmentation network, which contains pixel-level defect segmentation information.

[0126] Optionally, the interpretable segmentation network for constructing the multi-angle image can be obtained by training the model architecture (Grad-CAM++ integrated with ResNet50-ASPP) with weakly supervised labels.

[0127] Optionally, the calculation of the viewpoint homography matrix of the multi-angle image can be performed by feature point detection and matching.

[0128] S4. Based on the segmented multi-angle images, calculate the disintegration time and dynamic brittleness index of the initially defective tablet to determine the severity of the defect in the initially defective tablet.

[0129] The present invention calculates the disintegration time and dynamic brittleness index of the initially defective tablet based on the segmented multi-angle images, which can quantitatively assess the severity of the defect and improve the accuracy of subsequent parameter optimization.

[0130] Specifically, the step of calculating the disintegration time and dynamic fragility index of the initially defective tablet based on the segmented multi-angle images includes:

[0131] The defect volume of the preliminary defective tablet is calculated based on the segmented multi-angle image.

[0132] Based on the defect volume, the permeation path of the initially defective tablet is analyzed;

[0133] Based on the described penetration path, the initially defective tablets are classified into open defects and closed defects;

[0134] Analyze the defect cross-sectional area and total cross-sectional area of ​​the initially defective tablet;

[0135] Based on the defect cross-sectional area and the total cross-sectional area of ​​the tablet, the open-type defect disintegration time of the initially defective tablet is calculated;

[0136] The closed-type defect disintegration time of the initially defective tablet was calculated using the following formula:

[0137] ;

[0138] in, This indicates the disintegration time for a pre-existing defective tablet corresponding to a closed-type defect. This represents the baseline disintegration time of a pre-defective tablet compared to a defect-free tablet. The material-related constants of the initially defective tablet are represented. This indicates the defect volume of the initially defective tablet. The total volume of the tablets indicating initial defects. Represents an exponential function;

[0139] Based on the disintegration time limit of the open-type defect and the disintegration time limit of the closed-type defect, the disintegration time limit of the initially defective tablet is determined;

[0140] Based on the segmented multi-angle images, the radius of curvature of the defect tip of the initially defective tablet is analyzed to calculate the maximum stress of the initially defective tablet;

[0141] The dynamic brittleness index of the initially defective tablet is calculated based on the maximum stress.

[0142] The defect volume refers to the three-dimensional space occupied by the defect area identified by segmenting multi-angle images on the initially defective tablet. The penetration path refers to the channel through which liquid (such as a dissolution medium) can enter the tablet through the defect area. The open defect refers to a defect that is connected to the tablet surface, allowing external liquids to directly or indirectly enter the tablet. Examples include surface cracks and through-holes. The closed defect refers to a defect that is completely surrounded by tablet material, not connected to the surface, and prevents liquids from directly entering the tablet. Examples include internal microcavities and incompletely fused areas. The defect cross-sectional area refers to the area occupied by the open defect on a cross-section (usually the largest cross-section) perpendicular to the tablet's main axis. The total cross-sectional area of ​​the tablet refers to the total area of ​​the entire tablet on a cross-section perpendicular to the main axis. The open defect disintegration time refers to the time required for the tablet to completely disintegrate under specified conditions due to the presence of an open defect. The closed defect disintegration time refers to the effect of the presence of a closed defect on the tablet's disintegration time. The material-related constant refers to an empirical constant related to the tablet material properties (such as hardness, porosity, hydrophilicity, etc.). It reflects the inherent sensitivity of the material to defects. The radius of curvature at the defect tip refers to the radius of curvature at the tip of the defect (especially sharp cracks or notches). The maximum stress refers to the peak stress reached at the defect tip or other critical areas. The dynamic brittleness index is a quantitative indicator of the tablet's ability to resist breakage or wear under simulated dynamic stress (such as vibration or impact).

[0143] Optionally, the calculation of the defect volume of the preliminary defective tablet based on the segmented multi-angle image is obtained by performing three-dimensional reconstruction or voxel counting on the portion marked as defect in the segmented multi-angle image.

[0144] Optionally, analyzing the radius of curvature of the defect tip of the initially defective tablet to calculate the maximum stress of the initially defective tablet includes:

[0145] Determine the characteristic length and nominal stress of the initially defective tablet;

[0146] Based on the characteristic length, nominal stress, and radius of curvature of the defect tip, the maximum stress of the preliminary defective tablet is calculated using the following formula:

[0147] ;

[0148] in, This indicates the maximum stress of the initially defective tablet. The nominal stress of the initially defective tablet is indicated. The characteristic length of a pre-defective tablet. This indicates the radius of curvature of the defect tip of the initially defective tablet.

[0149] The characteristic length refers to a geometric parameter used to characterize the size of the defect. It can be automatically identified and measured from the segmented multi-angle image by edge detection. The maximum extension size of the defect in the tablet is half of the maximum extension size of the defect. The nominal stress refers to the macroscopic average stress that the tablet bears under specific working conditions when there are no defects. It can be measured by a pressure sensor.

[0150] Optionally, the dynamic friability index of the initially defective tablet can be calculated using the following formula:

[0151] ;

[0152] in, The dynamic fragility index indicates the initial defects of the tablet. Indicates the yield strength of a tablet with initial defects. This represents the correction factor. This indicates the maximum stress of a tablet with initial defects.

[0153] The yield strength refers to the critical stress value at which the tablet material begins to undergo plastic deformation. It is an inherent mechanical property of the material itself and can be determined through standard experiments. The correction factor refers to an empirical coefficient used to calibrate the deviation between theoretical calculations and actual experimental results. The correction factor in this invention can be 1.

[0154] Further, the calculation of the open-type defect disintegration time of the initially defective tablet based on the defect cross-sectional area and the total cross-sectional area of ​​the tablet includes:

[0155] Identify the liquid viscosity and concentration gradient of the dissolution medium corresponding to the initially defective tablet;

[0156] Analyze the drug diffusion coefficient of the initially defective tablet in the dissolution medium;

[0157] Based on the liquid viscosity, the concentration gradient, the drug diffusion coefficient, the defect cross-sectional area, and the total cross-sectional area of ​​the tablet, the disintegration time of the initially defective tablet corresponding to the open defect is calculated using the following formula:

[0158] ;

[0159] in, This indicates the disintegration time for an open-type defect tablet corresponding to a preliminary defect. Represents pi (π). Indicates the viscosity of the liquid. The tablet thickness indicates the initial defect of the tablet. Indicates the drug diffusion coefficient. Represents the concentration gradient. Indicates the cross-sectional area of ​​the defect. This indicates the total cross-sectional area of ​​the tablet.

[0160] The dissolution medium refers to a liquid used to simulate the drug's dissolution environment in vivo. The liquid viscosity refers to the degree of viscosity of the dissolution medium, i.e., its ability to resist flow. The higher the viscosity, the more difficult the liquid flow. The concentration gradient refers to the ratio of the concentration difference between the drug inside the tablet (high concentration) and the dissolution medium (low concentration), or the distance between the drug diffuses from the tablet surface to the dissolution medium away from the tablet. The drug diffusion coefficient is a measure of the ability of drug molecules in a pre-defective tablet to diffuse in a specific dissolution medium, representing the amount of drug molecules in a pre-defective tablet that passes through a unit area per unit time under a unit concentration gradient.

[0161] It should be explained that the defect severity of the preliminary defective tablets refers to the quantification of the severity of the defects in tablets that are initially identified as defective. Specifically, the defect severity can be calculated by defining the weights of the disintegration time and dynamic fragility index of the preliminary defective tablets, and then performing a weighted summation.

[0162] S5. Based on the severity of the defect, construct a defect cause analysis map of the preliminary defective tablet to generate equipment defect optimization parameters for the tablet production scenario, and execute intelligent equipment control for the tablet production scenario based on the equipment defect optimization parameters.

[0163] Based on the severity of the defects, this invention constructs a defect cause analysis map of the preliminary defective tablets, which can trace the possible root causes of the defects, thereby achieving precise optimization and control of equipment in tablet production scenarios.

[0164] Specifically, constructing the defect cause analysis map of the preliminary defective tablet based on the defect severity includes:

[0165] Obtain the defect-related process data of the preliminary defective tablet;

[0166] Based on the defect-related process data, a defect-process multidimensional feature vector is constructed for the preliminary defect tablet.

[0167] Calculate the causal correlation weights of the defect-process multidimensional feature vector;

[0168] Based on the causal correlation weights and the defect severity, a defect causal analysis map of the preliminary defective tablet is constructed.

[0169] The defect-related process data refers to the records of various parameters and conditions related to the manufacturing process when the initially defective tablets are produced, such as tableting machine pressure (kN), punch speed (mm / s), ambient temperature and humidity, and material moisture content. The defect-process multidimensional feature vector refers to the preprocessing (such as normalization, encoding, etc.) of the above-mentioned defect-related process data and organizing it into a structured vector. For example, the vector may include: raw material moisture content (dimensional 1), mixing time (dimensional 2), tableting pressure (dimensional 3), tablet hardness (dimensional 4), etc. The causal association weight refers to the quantitative index that measures the correlation strength between each process parameter (i.e., each dimension in the defect-process multidimensional feature vector) and the defect severity. The defect cause analysis map refers to a visual chart that intuitively shows how the defect severity of the initially defective tablets is related to different process factors.

[0170] Optionally, the defect cause analysis map of the preliminary defect tablet, constructed based on the cause correlation weight and the defect severity, can utilize a heatmap: with process parameters as rows / columns and cause correlation weights as color intensities, to visually display the correlation strength between each parameter and the defect.

[0171] Finally, this invention generates equipment defect optimization parameters for the tablet production scenario, and performs intelligent equipment control based on these parameters to achieve efficient and intelligent control of the tablet production equipment. The equipment defect optimization parameters refer to adjusted and optimized operating parameters set for production equipment (such as tablet presses, mixers, granulation equipment, etc.) to reduce or eliminate specific types of defects (such as disintegration, brittleness, and appearance defects) that occur during tablet production. Examples of such parameters include pressure, speed, filling depth, lubricant position, and temperature.

[0172] The experimental comparisons are as follows:

[0173] Experimental subject: Paliperidone sustained-release tablets;

[0174] Experimental groups: Traditional group, in which the traditional group underwent manual visual inspection (Method 1), where experienced quality inspectors inspected each tablet individually under standard lighting conditions according to SOP (Standard Operating Procedure); and Method 2, a single-modal machine vision system, which used a high-resolution industrial camera to capture images of the tablet surface and performed defect detection using preset image processing algorithms (such as threshold segmentation, edge detection, template matching, color analysis, etc.). Judgment criteria: The quality inspector made a final judgment on whether the tablet was qualified and the type of defect by combining the results of manual visual inspection and the results of the single-modal machine vision system.

[0175] Multimodal intelligent group: The equipment intelligent control method and system applied to tablet defect detection according to the above S1-S5 are used for detection. Judgment criteria: The system automatically outputs the defect type, location, and severity assessment results, and generates defect cause analysis map and equipment optimization parameter suggestions;

[0176] Quantification of experimental data:

[0177] Traditional Group: Number of defective paliperidone extended-release tablets detected: 80 (20 missed); 10 qualified tablets were misclassified as defective; Accuracy: (910 qualified correct + 80 defective correct) / 1000 = 99%; Precision: 80 / (80 + 10) = 88.9%; Recall: 80 / (80 + 20) = 80%; Single tablet detection time: 3 seconds (manual); Recall rate for internal defects such as disintegration abnormalities may only be 50%; Quantitative information such as defect volume and severity cannot be provided. Summary: Adjustments based on experience are uncertain.

[0178] Multimodal intelligent group: Number of defective paliperidone extended-release tablets detected: 95 (5 missed); 5 qualified tablets were misclassified as defective; Accuracy: (905 qualified correct + 95 defective correct) / 1000 = 99.5%; Precision: 95 / (95 + 5) = 95%; Recall: 95 / (95 + 5) = 95%; Single tablet detection time: 1.5 seconds; Recall rate for internal defects such as disintegration abnormalities may reach 90%; Quantitative information such as volume and severity (e.g., predicted disintegration time, fragility index) can be provided for all detected defects; Defect cause analysis map is generated, and the defect rate decreases by 15% in subsequent small-batch validation.

[0179] Therefore, compared with traditional inspection methods that rely on manual visual inspection and single-modal machine vision, the inspection scheme based on multimodal data fusion and intelligent analysis has significant advantages in defect detection of paliperidone extended-release tablets. It not only improves the accuracy, efficiency, and comprehensiveness of inspection, but more importantly, it can deeply analyze the causes of defects and provide data-driven intelligent decision support for optimizing the production process, thus helping to improve drug quality, stabilize production processes, and reduce costs.

[0180] First, by fusing multimodal data such as surface images, 3D morphology, and tablet compression acoustic signals, this system can acquire more comprehensive and in-depth product information than traditional single-modal detection. This greatly improves the accuracy and comprehensiveness of defect detection, enabling the identification of complex defects that were previously difficult to detect, such as internal microcracks and density inhomogeneities, effectively reducing the false negative rate and ensuring the quality and safety of pharmaceuticals. Second, the solution goes beyond defect identification; it extracts high-order features through a multimodal feature interaction model and uses a defect classification model for preliminary analysis, obtaining preliminary parameters including defect type and location. This lays the foundation for subsequent detailed analysis. Subsequently, through pixel-level segmentation of multi-angle images and calculation of key quality attributes (such as disintegration time and friability), The system can quantitatively assess the severity of defects, achieving a qualitative leap from "whether there is a defect" to "how severe the defect is," providing a more scientific basis for subsequent quality grading or handling. More innovative and valuable is its close integration of defect analysis with production process optimization. The causal analysis map built based on defect severity can trace the possible root causes of defects, whether they are raw material problems, formulation issues, or improper equipment parameter settings. The resulting optimized equipment defect parameters are then used to execute intelligent equipment control, forming a closed-loop system from detection to feedback and optimization. This not only improves the pass rate of single batches of products but also enhances the stability and consistency of the entire production process by continuously optimizing production equipment parameters, reducing the scrap rate. Therefore, this invention can improve the accuracy of tablet defect detection.

[0181] Example 2:

[0182] like Figure 2 The diagram shown is a functional block diagram of an intelligent control system for tablet defect detection according to the present invention.

[0183] The intelligent control system 200 for tablet defect detection described in this invention can be installed in an electronic device. Depending on the functions implemented, the intelligent control system for tablet defect detection may include a tablet data feature extraction module 201, a preliminary defect analysis module 202, a multi-angle image segmentation module 203, a defect severity analysis module 204, and an intelligent control module 205. The module described in this invention can also be referred to as a unit, which refers to a series of computer program segments that can be executed by the processor of an electronic device and perform a fixed function, stored in the memory of the electronic device.

[0184] In this embodiment of the invention, the functions of each module / unit are as follows:

[0185] The tablet data feature extraction module 201 is used to collect tablet data of tablets to be tested in a tablet production scenario. The tablet data includes surface images, three-dimensional morphology data and acoustic signals of the tablet compression process. The module extracts the data features of the tablet data to obtain tablet data features.

[0186] The preliminary defect analysis module 202 is used to analyze the high-order feature vector of the tablet data features using the multimodal feature interaction model in the pre-constructed defect screening network, and based on the high-order feature vector, analyze the defect parameters of the tablet to be tested using the defect classification model in the defect screening network, wherein the defect parameters include preliminary defect type label and preliminary defect location.

[0187] The multi-angle image segmentation module 203 is used to mark the defect parameters corresponding to the preliminary defect tablets in the tablets to be tested, and to acquire multi-angle images of the preliminary defect tablets. Based on the defect parameters, the multi-angle images are segmented at the pixel level to obtain segmented multi-angle images.

[0188] The defect severity analysis module 204 is used to calculate the disintegration time and dynamic fragility index of the initially defective tablet based on the segmented multi-angle images, so as to determine the defect severity of the initially defective tablet.

[0189] The intelligent equipment control module 205 is used to construct a defect cause analysis map of the preliminary defective tablet based on the defect severity, so as to generate equipment defect optimization parameters for the tablet production scenario, and execute intelligent equipment control for the tablet production scenario based on the equipment defect optimization parameters.

[0190] In detail, the modules in the intelligent control system 200 for tablet defect detection described in this embodiment of the invention employ the same methods as described above during use. Figure 1 The same technical means are used in the intelligent control method for tablet defect detection described in the article, and can produce the same technical effect, so it will not be repeated here.

[0191] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0192] Finally, 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.

Claims

1. A method for intelligent control of equipment applied to tablet defect detection, characterized in that, The method includes: Tablet data of tablets to be tested in a tablet production scenario is collected, wherein the tablet data includes surface images, three-dimensional morphology data and acoustic signals of the tableting process, and the data features of the tablet data are extracted to obtain tablet data features; The high-order feature vector of the tablet data features is analyzed using a multimodal feature interaction model in a pre-constructed defect screening network. Based on the high-order feature vector, the defect parameters of the tablet to be tested are analyzed using a defect classification model in the defect screening network. The defect parameters include preliminary defect type labels and preliminary defect locations. The analysis of the high-order feature vector of the tablet data features using the multimodal feature interaction model in the pre-constructed defect screening network includes: calculating the gating weights of the time-domain-frequency domain joint features in the tablet data features; calculating the attention weights of the multi-scale feature map and point cloud global features in the tablet data features using the attention layer in the multimodal feature interaction model based on the gating weights; aggregating the tablet data features using the aggregation layer in the multimodal feature interaction model according to the attention weights to obtain aggregated features; and reducing the dimensionality of the aggregated features using the feature dimensionality reduction layer in the multimodal feature interaction model to obtain the high-order feature vector of the tablet data features. The defect parameters are labeled to correspond to the preliminary defective tablets in the tablets to be tested, and multi-angle images of the preliminary defective tablets are acquired. Based on the defect parameters, the multi-angle images are segmented at the pixel level to obtain segmented multi-angle images. Based on the segmented multi-angle images, the disintegration time and dynamic fragility index of the initially defective tablet are calculated to determine the severity of the defect in the initially defective tablet. Based on the severity of the defect, a defect cause analysis map of the preliminary defective tablet is constructed to generate equipment defect optimization parameters for the tablet production scenario. Based on the equipment defect optimization parameters, intelligent equipment control for the tablet production scenario is executed.

2. The intelligent control method for tablet defect detection as described in claim 1, characterized in that, The extraction of data features from the tablet data to obtain tablet data features includes: Calculate the extrinsic matrix of the surface image and three-dimensional morphology data in the tablet data; Based on the extrinsic parameter matrix, a mapping three-dimensional point cloud coordinate system is established between the surface image and the three-dimensional topography data; Establish a sound source-image region mapping table for the acoustic signals of the tableting process in the surface image and the tablet data; Based on the mapped 3D point cloud coordinate system and the sound source-image region mapping table, the tablet data is spatially registered to obtain registered tablet data, wherein the registered tablet data includes: registered surface image, registered 3D morphology data and registered acoustic signal of tableting process; Multi-scale feature maps, point cloud global features, and time-frequency domain joint features of the registration surface image, registration three-dimensional topography data, and acoustic signals of the registration pressing process are extracted respectively. Based on the multi-scale feature map, point cloud global features, and time-frequency domain joint features, the tablet data features of the tablet data are determined.

3. The intelligent control method for tablet defect detection as described in claim 2, characterized in that, The extraction of multi-scale feature maps, point cloud global features, and time-domain-frequency domain joint features from the registration surface image, registration 3D topography data, and acoustic signals from the registration pressing process includes: The registered surface image is downsampled to obtain a feature map sequence; The Scharr gradient operator of the feature map sequence is calculated to extract multi-scale feature maps of the sampled and registered surface image; The registered 3D topography data is divided into a voxel mesh; Identify the principal curvature direction of the tablet in the voxel mesh to perform voxel reconstruction on the voxel mesh, thereby obtaining a reconstructed voxel mesh; The reconstructed voxel mesh is subjected to density-curvature joint filtering to obtain a filtered reconstructed voxel mesh; Calculate the seven-dimensional feature vector of the filtered and reconstructed voxel mesh to determine the global point cloud features of the filtered and reconstructed voxel mesh; Analyze the Mel spectrum of the acoustic signal during the registration and tableting process; Based on the Mel spectrum, identify the time-frequency joint characteristics of the acoustic signal during the registration and tableting process.

4. The intelligent control method for tablet defect detection as described in claim 3, characterized in that, The analysis of the Mel spectrum of the acoustic signal during the registration and tableting process includes: Analyze the minimum and maximum frequencies of the acoustic signals during the registration and tableting process; Analyze the resonant peak shift coefficient and nonlinear intensity coefficient of the acoustic signal during the registration and compression process; Based on the minimum frequency of the frequency band, the maximum frequency of the frequency band, the resonant peak shift coefficient, and the nonlinear intensity coefficient, the filter center frequency of the acoustic signal during the registration and pressing process is calculated using the following formula: ； in, The acoustic signal representing the registration and tableting process. The center frequency of each filter. Indicates the minimum frequency of the frequency band. Indicates the maximum frequency of the frequency band. The acoustic signal representing the registration and tableting process. One filter, This indicates the number of filters used to transmit the acoustic signals during the registration and tableting process. This represents the resonance peak shift coefficient. Represents the nonlinear strength coefficient. Represents the arctangent function; Based on the center frequency of the filter, the Mel spectrum of the acoustic signal during the registration and compression process is analyzed.

5. The intelligent control method for tablet defect detection as described in claim 4, characterized in that, The step of performing pixel-level segmentation on the multi-angle image based on the defect parameters to obtain a segmented multi-angle image includes: Based on the defect parameters, weakly supervised labels are generated for the multi-angle images to construct an interpretable segmentation network for the multi-angle images. Calculate the viewpoint homography matrix of the multi-angle image; Based on the interview homography matrix, the multi-angle images are spatially aligned to obtain aligned multi-angle images; The illustrative segmentation network is used to perform pixel-level segmentation on the aligned multi-angle image to obtain the segmented multi-angle image.

6. The intelligent control method for tablet defect detection as described in claim 5, characterized in that, The step of calculating the disintegration time and dynamic fragility index of the initially defective tablet based on the segmented multi-angle images includes: The defect volume of the preliminary defective tablet is calculated based on the segmented multi-angle image. Based on the defect volume, the permeation path of the initially defective tablet is analyzed; Based on the described penetration path, the initially defective tablets are classified into open defects and closed defects; Analyze the defect cross-sectional area and total cross-sectional area of ​​the initially defective tablet; Based on the defect cross-sectional area and the total cross-sectional area of ​​the tablet, the open-type defect disintegration time of the initially defective tablet is calculated; Calculate the closed-type defect disintegration time of the initially defective tablet; Based on the disintegration time limit of the open-type defect and the disintegration time limit of the closed-type defect, the disintegration time limit of the initially defective tablet is determined; Based on the segmented multi-angle images, the radius of curvature of the defect tip of the initially defective tablet is analyzed to calculate the maximum stress of the initially defective tablet; The dynamic brittleness index of the initially defective tablet is calculated based on the maximum stress.

7. The intelligent control method for tablet defect detection as described in claim 6, characterized in that, The calculation of the open-type defect disintegration time of the initially defective tablet based on the defect cross-sectional area and the total cross-sectional area of ​​the tablet includes: Identify the liquid viscosity and concentration gradient of the dissolution medium corresponding to the initially defective tablet; Analyze the drug diffusion coefficient of the initially defective tablet in the dissolution medium; Based on the liquid viscosity, the concentration gradient, the drug diffusion coefficient, the defect cross-sectional area, and the total cross-sectional area of ​​the tablet, the disintegration time of the open-type defect corresponding to the initially defective tablet is calculated.

8. The intelligent control method for tablet defect detection as described in claim 7, characterized in that, The process of constructing a preliminary defect cause analysis map of the defective tablet based on the defect severity includes: Obtain the defect-related process data of the preliminary defective tablet; Based on the defect-related process data, a defect-process multidimensional feature vector is constructed for the preliminary defect tablet. Calculate the causal correlation weights of the defect-process multidimensional feature vector; Based on the causal correlation weights and the defect severity, a defect causal analysis map of the preliminary defective tablet is constructed.

9. An intelligent control system for tablet defect detection, characterized in that, The system includes: The tablet data feature extraction module is used to collect tablet data of tablets to be tested in a tablet production scenario. The tablet data includes surface images, three-dimensional morphology data and acoustic signals of the tableting process. The module extracts the data features of the tablet data to obtain tablet data features. The preliminary defect analysis module is used to analyze the high-order feature vector of the tablet data features using a pre-constructed multimodal feature interaction model in the defect screening network, and based on the high-order feature vector, analyze the defect parameters of the tablet to be tested using a defect classification model in the defect screening network. The defect parameters include a preliminary defect type label and a preliminary defect location. The step of analyzing the high-order feature vector of the tablet data features using the pre-constructed multimodal feature interaction model in the defect screening network includes: calculating the gating weights of the time-domain-frequency domain joint features in the tablet data features; based on the gating weights, calculating the attention weights of the multi-scale feature map and point cloud global features in the tablet data features using the attention layer in the multimodal feature interaction model; aggregating the tablet data features using the aggregation layer in the multimodal feature interaction model according to the attention weights to obtain aggregated features; and reducing the dimensionality of the aggregated features using the feature dimensionality reduction layer in the multimodal feature interaction model to obtain the high-order feature vector of the tablet data features. A multi-angle image segmentation module is used to mark the defect parameters corresponding to the preliminary defect tablets in the tablets to be detected, and to acquire multi-angle images of the preliminary defect tablets. Based on the defect parameters, the multi-angle images are segmented at the pixel level to obtain segmented multi-angle images. The defect severity analysis module is used to calculate the disintegration time and dynamic fragility index of the initially defective tablet based on the segmented multi-angle images to determine the defect severity of the initially defective tablet; the equipment intelligent control module is used to construct a defect cause analysis map of the initially defective tablet based on the defect severity to generate equipment defect optimization parameters for the tablet production scenario, and execute equipment intelligent control for the tablet production scenario based on the equipment defect optimization parameters.

Images