Ceramic slurry production monitoring method for multilayer ceramic capacitor

Abstract

The invention provides a ceramic slurry production monitoring method for a multilayer ceramic capacitor, and relates to the technical field of production monitoring. The ceramic slurry production monitoring method comprises the steps from S01 to S07. S01, obtaining a ceramic slurry sample. And S02, detecting bubble information of the ceramic slurry sample by using an ultrasonic detector. And S03, measuring the viscosity of the ceramic slurry sample at a preset temperature and a preset rotating speed by using a rotary viscometer. And S04, drying the ceramic slurry sample with a preset mass, and calculating the mass percentage of solid residues after drying to obtain the solid content of the ceramic slurry sample. And S05, preparing the ceramic slurry sample into a dry smear, and obtaining a microscope image of the smear. And S06, identifying particle information in the microscope image through the particle identification model. Wherein the particle information comprises the particle type and the particle size of each particle. And S07, according to the bubble information, the viscosity, the solid content, the particle type and the particle size, judging whether the ceramic slurry is qualified or not by referring to a preset judgment standard.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of production monitoring, in particular to a porcelain slurry production monitoring method for multilayer ceramic capacitors. BACKGROUND

[0002] In the production process of multilayer ceramic capacitors, the quality of porcelain slurry is a key factor affecting the performance and reliability of capacitors.

[0003] The production of porcelain slurry involves various complex process parameters, such as raw material ratio, stirring time, temperature, etc., which together determine the final quality of the porcelain slurry. However, the existing porcelain slurry quality detection methods have significant shortcomings in particle size identification.

[0004] Particle size is one of the important indicators to measure the quality of porcelain slurry, and accurate particle size distribution is crucial to ensure the uniformity and stability of porcelain slurry. However, current technical means in identifying particle size are often limited by factors such as detection equipment precision, image processing algorithm complexity, and particle morphology diversity, resulting in inaccurate particle size measurement results. This inaccurate particle size identification makes it difficult to accurately judge whether the quality of porcelain slurry is qualified, thereby affecting the overall production quality and efficiency of multilayer ceramic capacitors.

[0005] In addition, when the quality of porcelain slurry is unqualified, the existing technology lacks effective means to quickly locate the specific parameters causing the problem, making quality control and optimization in the production process face great challenges. SUMMARY

[0006] The present application provides a porcelain slurry production monitoring method for multilayer ceramic capacitors to improve at least one of the above technical problems. The porcelain slurry production monitoring method comprises steps S01 to S07.

[0007] S01, obtaining a porcelain slurry sample.

[0008] S02, using an ultrasonic detector to detect the bubble information of the porcelain slurry sample.

[0009]

[0009] S03, using a rotary viscometer to measure the viscosity of the porcelain slurry sample at a preset temperature and a preset rotation speed.

[0010] S04, drying a porcelain slurry sample of a preset mass, and calculating the mass percentage of solid residues after drying to obtain the solid content of the porcelain slurry sample.

[0011] S05, preparing a dry smear of the porcelain slurry sample, and obtaining a microscope image of the smear.

[0012] S06, identifying particle information in the microscope image through a particle identification model. The particle information includes the particle type and particle size of each particle.

[0013] S07、According to the bubble information, the viscosity, the solid content, the particle type and the particle size, referring to a preset judgment standard, judging whether the porcelain slurry is qualified.

[0014] In an optional embodiment, step S06 specifically comprises steps S061 to S065.

[0015] S061, preprocessing the microscope image.

[0016] S062, obtaining mask information of each particle in the preprocessed microscope image through a particle mask model.

[0017] S063, according to the mask information, performing particle separation and morphological operation to obtain independent particle contours.

[0018] S064, according to the independent particle contours, obtaining the particle size of each particle.

[0019] S065, according to the mask information, the particle size and the preprocessed microscope image, extracting multiple features of each particle and classifying to obtain the particle type of each particle.

[0020] In an optional embodiment, the particle mask model comprises a first input layer, a self-defined backbone network connected to the first input layer, a region proposal network connected to the self-defined backbone network, a detection branch connected to the self-defined backbone network and the region proposal network, a mask branch connected to the self-defined backbone network and the detection branch, and a first output layer connected to the mask branch.

[0021] In an optional embodiment, step S062 specifically comprises steps S0621 to S0625.

[0022] S0621, inputting the preprocessed microscope image into the self-defined backbone network through the first input layer to extract feature maps.

[0023] S0622, inputting the feature maps extracted by the self-defined backbone network into the region proposal network to generate candidate regions that may contain particles on the feature maps.

[0024] S0623, according to the candidate regions, judging whether the candidate regions contain particles through the detection branch.

[0025] S0624, according to the judgment result of the detection branch, generating a binary mask for the candidate regions containing particles using the mask branch to segment and obtain the specific contours of the particles in the image, and obtaining the mask information of each particle.

[0026] S0625. Output the mask information and position information of the particles through the first output layer.

[0027] In an optional implementation, step S063 specifically includes steps S0631 to S0636.

[0028] S0631. Based on the mask information, perform morphological opening operation. First, detect whether there are voids inside the particles. If there are voids, use morphological closing operation to fill them. Then, use erosion operation to remove noise from the mask edge. Then, use dilation operation to restore the shape of the particles. Finally, delete particles with a pixel count less than a first preset pixel threshold to obtain the morphologically processed mask information.

[0029] S0632. Perform distance transformation on the morphologically processed mask information, calculate the Euclidean distance from each foreground pixel to the nearest background pixel, and obtain the distance-transformed grayscale image. The value of each pixel in the grayscale image represents its distance to the background.

[0030] S0633. Based on the grayscale image, the center point of each particle is found using a local maximum detection algorithm. Then, a blank image of the same size as the input image is created, and the center points are marked in the blank image to generate a particle-marked image. Each center point in the particle-marked image represents an independent particle.

[0031] S0634. Based on the distance-transformed grayscale image and the particle marker image, the watershed algorithm is used to segment the grayscale image to separate the particles adhering to each other in the mask image and obtain the contour of each individual particle.

[0032] S0635. Detect whether the outlines of two adjacent individual particles are connected.

[0033] S0636. For the contours of two connected individual particles, calculate the edge smoothness of the independent contours of the two contours, and the edge smoothness of the merged contour of the outer contour after merging the two contours. When the edge smoothness of the merged contour is less than the edge smoothness of the two independent contours, merge the two connected contours. Finally, delete contours with a pixel count less than a first preset pixel threshold to obtain independent particle contours.

[0034] The calculation model for edge smoothness is as follows:

[0035] .

[0036] .

[0037] .

[0038] .

[0039] In the formula, For edge smoothness, The number of points on the outline, Numbering the points on the outline For the curvature change between adjacent points, For the first The curvature of each point, For the first The curvature of each point, For the first The angle between vectors at each point, For the first The point and the first Distance between points, For the first The point and the first Distance between points, For the first The coordinates of each point For the first The coordinates of each point For the first The coordinates of each point For the first The point points to the first Vector of points, For the first The point points to the first A vector of points.

[0040] In an optional implementation, step S064 specifically includes steps S0641 to S0643.

[0041] S0641. Based on the independent particle profile, determine the length of the major axis and the length of the minor axis using the minimum bounding rectangle algorithm, and calculate the aspect ratio of the particle.

[0042] S0642. Based on the magnification of the microscope image and the individual particle outlines, obtain the area and perimeter of each particle's outline. The area is calculated using a pixel counting method. The perimeter is calculated using a pixel spacing method.

[0043] S0643. Based on the area and the perimeter, and considering the shape factor, obtain the particle size. The calculation model for the particle size is as follows:

[0044] .

[0045] .

[0046] In the formula, a particle size, an area, a perimeter, a shape factor.

[0047] In an optional embodiment, the step S065 specifically comprises steps S0651 to S0652.

[0048] S0651, extracting texture features from the mask information and the pre-processed microscopic image through a gray level co-occurrence matrix. The texture features include contrast, correlation, and energy.

[0049] S0652, inputting the particle contour, aspect ratio, area, perimeter, shape factor, particle size, and the texture features of the particle into a particle classification model to obtain the particle type. The particle type includes ceramic powder particles, porcelain slurry agglomerated particles, residual solid particles, and porcelain slurry impurity particles.

[0050] In an optional embodiment, the preset judgment criteria include:

[0051] Porcelain slurry bubble threshold: no more than 10 bubbles per cubic centimeter, and the maximum bubble diameter is no more than 5 μm.

[0052] Viscosity threshold: 1000 to 5000 mPa.s.

[0053] Solid content threshold: 55±3%.

[0054] Particle size threshold: average particle size 100 nm to 300 nm, median particle size D50 no more than 200 nm, and 90% of the particles D90 no more than 300 nm.

[0055] Binder agglomerated particle threshold: no visible agglomerated block, and no agglomerated particle larger than 500 nm at any position.

[0056] Residual solid particle threshold: no more than 3 particles per square centimeter of surface, and the residual solid particle size is less than 500 nm.

[0057] Porcelain slurry impurity particle threshold: no more than 3 particles per square centimeter of surface, and the residual solid particle size is less than 500 nm.

[0058] In an optional embodiment, the porcelain slurry production monitoring method further comprises steps S08 to S10.

[0059] S08, when it is judged that the porcelain slurry is unqualified, obtaining the related parameters of the porcelain slurry. The related parameters include production parameters that can be controlled in the production process and characteristic parameters of the porcelain slurry after production. The production parameters include production personnel information, production machine information, production material information, raw material ratio information, stirring speed, stirring time, temperature in the batching room, humidity in the batching room, porcelain slurry container temperature, and porcelain slurry container humidity. The characteristic parameters include bubble information, viscosity, solid content, and the particle information.

[0060] S09, inputting the related parameters of the unqualified porcelain slurry into a defect cause analysis model to obtain a set of unqualified production parameters.

[0061] S10, inputting the related parameters and the set of unqualified production parameters into a defect cause solving model to obtain suggested values of the unqualified production parameters.

[0062] By adopting the above technical solution, the following technical effects can be achieved.

[0063] The application can significantly improve the judgment accuracy of the porcelain slurry quality by optimizing the particle recognition algorithm and multi-parameter comprehensive analysis. First, the particle mask model constructed based on deep learning can accurately segment the particles in the microscope image, effectively solve the problems of particle adhesion and edge noise by combining morphological operations and watershed algorithm, and ensure the accurate extraction of the independent particle contour. Second, the particle size calculation model with dynamic correction of the shape factor is introduced, which comprehensively considers the actual morphological characteristics of the particles, and greatly improves the reliability of the particle size data. In addition, the texture features are extracted by the gray level co-occurrence matrix, and combined with multi-dimensional data such as particle contour and aspect ratio, the efficient classification of ceramic powder particles, agglomerated particles, residual solid particles and impurity particles is realized.

[0064] Finally, the systematic judgment standard integrates multi-parameter data such as bubbles, viscosity, solid content, particle type and particle size to form a comprehensive quality evaluation system, avoiding the one-sidedness of a single index. This method not only improves the automation level of detection and reduces the error of human intervention, but also quickly locates the production parameter abnormalities of unqualified porcelain slurry, provides reliable basis for process optimization, and thus guarantees the high performance and stability of the multilayer ceramic capacitor. BRIEF DESCRIPTION OF DRAWINGS

[0065] In order to more clearly illustrate the technical solutions of the present application, the following will briefly introduce the drawings needed to be used in the specific embodiments of the present application. It should be understood that the following drawings only show some specific embodiments of the present application, and therefore should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can also be obtained without creative labor.

[0066] Figure 1A flowchart of a porcelain slurry production monitoring method of a multilayer ceramic capacitor.

[0067] Figure 2 A flowchart of particle information recognition.

[0068] Figure 3 A network structure diagram of a particle mask model containing detailed information of a custom backbone network.

[0069] Figure 4 A network structure diagram of a particle mask model containing detailed information of a region proposal network.

[0070] Figure 5 A network structure diagram of a particle mask model containing detailed information of a detection branch / mask branch.

[0071] Figure 6 A network structure diagram of a particle classification model.

[0072] Figure 7 A network structure diagram of a defect cause analysis model.

[0073] Figure 8 A network structure diagram of a defect cause solving model. DETAILED DESCRIPTION

[0074] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0075] Please refer to Figures 1 to 8 The first embodiment of the present application provides a porcelain slurry production monitoring method of a multilayer ceramic capacitor, which can be executed by a porcelain slurry production monitoring method device (hereinafter referred to as a porcelain slurry production monitoring device). In particular, it is executed by one or more processors in the porcelain slurry production monitoring device to realize steps S01 to S07.

[0076] S01, obtaining a porcelain slurry sample.

[0077] It can be understood that the porcelain slurry production monitoring device can be a portable notebook computer, a desktop computer, a server, a smart phone or a tablet computer, etc. electronic equipment with computing performance.

[0078] Specifically, the ceramic slurry is a ceramic slurry used to produce ceramic capacitors. After the ceramic slurry is stirred, a sample collector is used to collect a suitable amount of sample at different positions and depths. During the sampling process, attention should be paid to avoid introducing impurities or air bubbles to ensure the purity of the sample. At the same time, the amount of sample should be sufficient for subsequent detection, and generally at least 3 or more parallel samples are collected per batch. After sampling, seal and store in time to prevent evaporation and contamination.

[0079] S02, using an ultrasonic detector, detecting the bubble information of the ceramic slurry sample.

[0080] The ultrasonic detection technology can be used to detect the bubble information in the ceramic slurry sample. The basic principle is that when ultrasonic waves propagate in a liquid, they will produce reflection and scattering phenomena when encountering bubbles. By analyzing the intensity, frequency and time difference of the reflection and scattering signals, the number, size and distribution of the bubbles can be obtained.

[0081] S03, using a rotary viscometer to measure the viscosity of the ceramic slurry sample at a predetermined temperature and a predetermined rotational speed.

[0082] The rotary viscometer is a commonly used instrument for measuring the viscosity of the ceramic slurry sample, and its basic structure includes a coaxial rotor and an outer cylinder. By controlling the rotational speed of the rotor and measuring the required torque, the viscosity value of the ceramic slurry at a given shear rate can be calculated.

[0083] S04, taking a predetermined mass of ceramic slurry sample for drying, and calculating the mass percentage of solid residue after drying to obtain the solid content of the ceramic slurry sample.

[0084] In order to obtain the solid content of the ceramic slurry sample, a predetermined mass of ceramic slurry sample needs to be dried. By calculating the percentage of mass before and after drying, the solid content of the ceramic slurry sample can be obtained. The solid content calculation formula is: solid content (%) = mass of solid residue after drying / mass of original ceramic slurry sample x 100%.

[0085] The solid content is an important parameter to characterize the concentration and composition of the ceramic slurry. If the solid content is too low, it means that the solid powder particle content in the ceramic slurry is insufficient, the slurry is too dilute, and the coating and sintering performance may be affected. If the solid content is too high, it means that the powder content is too high, the slurry viscosity is large, and it is not conducive to coating and forming. Therefore, measuring the solid content can monitor whether the formula of the ceramic slurry is correct and the concentration is appropriate. At the same time, the fluctuation of the solid content also reflects the stability of the process of feeding, mixing and dispersing in the preparation of the ceramic slurry, and is a means of quality control.

[0086] S05, preparing the ceramic slurry sample into a dry coating, and obtaining a microscope image of the coating.

[0087] Specifically, the microscope image is taken by a scanning electron microscope (SEM). The porcelain slurry sample is prepared into a dry smear, and a microscope image of the smear is obtained. The taken microscope image is saved as a digital image file to provide original image data for subsequent particle identification analysis.

[0088] S06, identifying particle information in the microscope image by a particle identification model. The particle information includes particle type and particle size of each particle.

[0089] The acquired microscope image is analyzed by using a particle identification model based on deep learning. The model first pre-processes the original image, including grayscale, denoising and other operations, to improve the image quality and the clarity of the particle outline. Then, the model extracts the features of the image through a self-defined backbone network, and generates candidate boxes that may contain particle targets by using a region proposal network. Next, the candidate boxes are further screened by a detection branch to accurately determine whether the candidate boxes contain particles. At the same time, the mask branch of the model can accurately segment the pixel-level mask of each particle to obtain the particle outline. Then, the equivalent particle size is calculated according to the outline. Finally, the particle classification model combined with convolutional neural network and fully connected neural network is used to accurately identify the type of the particle according to the characteristics of the particle, thereby providing key parameter basis for subsequent porcelain slurry quality judgment.

[0090] On the basis of the above embodiment, in an optional embodiment of the present application, as shown in Figure 2 Step S06 specifically includes steps S061 to S065.

[0091] S061, pre-processing the microscope image.

[0092] When pre-processing the microscope image, first, the image is denoised to remove random noise in the image by using Gaussian filtering and to retain the edge information of the particles. Next, the image is subjected to histogram equalization to enhance the contrast of the image and make the particles and the background more distinguishable. Finally, the image is converted into a grayscale image to reduce the calculation complexity while retaining the morphological information of the particles.

[0093] S062, obtaining mask information of each particle in the pre-processed microscope image by a particle mask model. Preferably, the particle mask model includes a first input layer, a self-defined backbone network connected to the first input layer, a region proposal network connected to the self-defined backbone network, a detection branch connected to the self-defined backbone network and the region proposal network, a mask branch connected to the self-defined backbone network and the detection branch, and a first output layer connected to the mask branch.

[0094] The step S062 specifically includes steps S0621 to S0625.

[0095] S0621. The preprocessed microscope image is input into a custom backbone network through the first input layer to extract feature maps. Specifically, when obtaining the mask information of each particle in the preprocessed microscope image through the particle mask model, the preprocessed image is first input into the custom backbone network. The custom backbone network is an improvement based on a mask-based region-based convolutional neural network, which can extract multi-level features from the image.

[0096] S0622. The feature map extracted by the custom backbone network is input into the region proposal network to generate candidate regions that may contain particles on the feature map. Specifically, the region proposal network uses an anchoring mechanism to generate multiple candidate regions of different scales and aspect ratios on the feature map. For example, for a feature map, the region proposal network may generate 1000 candidate regions, each containing a rectangular box that may contain particles.

[0097] S0623. Based on the candidate regions, determine whether the candidate regions contain particles using a detection branch. Specifically, based on the candidate regions, determine whether the candidate regions contain particles using a detection branch. The detection branch uses a binary classifier to classify each candidate region and determine whether it contains particles. For example, for 1000 candidate regions, the detection branch may determine that 200 candidate regions contain particles, and the remaining 800 candidate regions do not contain particles.

[0098] S0624. Based on the judgment result of the detection branch, a mask branch is used to generate a binary mask for the candidate region containing particles, in order to segment and obtain the specific contour of the particles in the image, and obtain the mask information of each particle. Specifically, the mask branch adopts a convolutional network structure to generate a binary mask for each candidate region containing particles, representing the specific contour of the particle. For example, for a candidate region containing particles, the mask branch generates a binary mask of the same size as the candidate region, where 1 represents the particle region and 0 represents the background region.

[0099] S0625. Output the mask information and position information of the particles through the first output layer. Specifically, the first output layer outputs the mask information and position information of the particles. The first output layer integrates the binary mask generated by the mask branch and the candidate region information generated by the detection branch, and outputs the mask information and position information of each particle.

[0100] S063. Based on the mask information, perform particle separation and morphological operations to obtain independent particle profiles. Preferably, step S063 specifically includes steps S0631 to S0636.

[0101] S0631、According to the mask information, morphological opening operation is performed to detect whether there is a hollow region in the particle, if there is a hollow, morphological closing operation is used to fill it, then noise on the edge of the mask is removed by using erosion operation, and then the shape of the particle is restored by using dilation operation, and finally the particles with a pixel number less than a first preset pixel threshold are deleted to obtain the mask information after morphological processing. Specifically, for a mask information, after morphological operation, the hollow in the particle is filled, the edge noise is removed, and the shape of the particle is restored.

[0102] S0632, The mask information after morphological processing is subjected to distance transformation, and the Euclidean distance of each foreground pixel to the nearest background pixel is calculated to obtain a gray scale image after distance transformation. Wherein, the value of each pixel of the gray scale image represents the distance to the background. Specifically, for a mask information, in the gray scale image after distance transformation, the value of each pixel represents the distance to the background, the farther the distance, the larger the pixel value, and the closer the distance, the smaller the pixel value.

[0103] S0633, According to the gray scale image, the center point of each particle is found using a local maximum value detection algorithm, then a blank image with the same size as the input image is created, and the center point is marked in the blank image to generate a particle marker image. Wherein, each center point in the particle marker image represents an independent particle. Specifically, for a gray scale image, the local maximum value detection algorithm may find 100 center points, each of which represents an independent particle.

[0104] S0634, According to the gray scale image after distance transformation and the particle marker image, the gray scale image is segmented to separate the adhered particles in the mask image using a watershed algorithm to obtain the contour of each single particle. Specifically, for a gray scale image and a particle marker image, the watershed algorithm divides the image into multiple single particle contours, each of which represents an independent particle.

[0105] S0635, Detect whether the contours of two adjacent single particles are connected.

[0106] S0636, For the contours of two single particles connected, the edge smoothness of the independent contours of the two contours and the edge smoothness of the merged contour of the outer contour after merging of the two contours are calculated respectively. When the edge smoothness of the merged contour is less than the edge smoothness of the two independent contours, the two connected contours are merged. Finally, the contours with a pixel number less than a first preset pixel threshold are deleted to obtain the independent particle contour. Specifically, for two connected contours, if the edge smoothness of the merged contour is less than the edge smoothness of the independent contour, the two contours are merged into one contour.

[0107] The calculation model for edge smoothness is as follows: .

[0108] .

[0109] .

[0110] .

[0111] In the formula, For edge smoothness, The number of points on the outline, Numbering the points on the outline For the curvature change between adjacent points, For the first The curvature of each point, For the first The curvature of each point, For the first The angle between vectors at each point, For the first The point and the first Distance between points, For the first The point and the first Distance between points, For the first The coordinates of each point For the first The coordinates of each point For the first The coordinates of each point For the first The point points to the first Vector of points, For the first The point points to the first A vector of points.

[0112] S064. Obtain the particle size of each particle based on the independent particle profile. Preferably, step S064 specifically includes steps S0641 to S0643.

[0113] S0641. Based on the independent particle profile, determine the length of the major axis and the length of the minor axis using the minimum bounding rectangle algorithm, and calculate the aspect ratio of the particle.

[0114] S0642. Based on the magnification of the microscope image and the individual particle outlines, obtain the area and perimeter of each particle's outline. The area is calculated using a pixel counting method. The perimeter is calculated using a pixel spacing method.

[0115] S0643. Based on the area and the perimeter, and considering the shape factor, obtain the particle size. The calculation model for the particle size is as follows: .

[0116] .

[0117] In the formula, The particle size, For area, For perimeter, This is the shape factor.

[0118] For example, for a single particle profile, the minimum bounding rectangle algorithm determines the major axis to be 50 pixels, the minor axis to be 30 pixels, and the aspect ratio to be 1.67. Based on the magnification and the profile, the calculated area is 1500 pixels and the perimeter is 200 pixels. Considering the shape factor, the calculated particle size is 25 micrometers.

[0119] S065. Based on the mask information, the particle size, and the preprocessed microscopic image, extract multiple features from each particle and classify them to obtain the particle type of each particle. The particle types include: ceramic powder particles, ceramic slurry agglomerated particles, residual solid particles, and ceramic slurry impurity particles. Preferably, step S065 specifically includes steps S0651 to S0652.

[0120] S0651. Based on the mask information and the preprocessed microscopic image, extract texture features using a gray-level co-occurrence matrix. The texture features include contrast, correlation, and energy.

[0121] S0652. Input the particle outline, aspect ratio, area, perimeter, shape factor, particle diameter, and texture features into the particle classification model to obtain the particle type.

[0122] For example, for a single particle, the contrast extracted using the gray-level co-occurrence matrix is ​​0.5, the correlation is 0.8, and the energy is 0.7. Inputting the particle's outline, aspect ratio, area, perimeter, shape factor, particle size, and texture features into the particle classification model yields a classification result of ceramic powder particles.

[0123] S07. Based on the bubble information, viscosity, solid content, particle type, and particle size, and referring to the preset judgment criteria, determine whether the porcelain slurry is qualified.

[0124] Preferably, to determine whether a porcelain slurry is qualified, multiple parameters such as bubble information, viscosity, solid content, particle type, and particle size need to be considered comprehensively. Preset judgment criteria include:

[0125] Porcelain slurry bubble threshold: no more than 10 bubbles per cubic centimeter, and the maximum bubble diameter is not more than 5 μm. Specifically, the bubble is one of the important factors affecting the quality of the porcelain slurry. Too many bubbles or too large size will cause the porcelain body to have internal hole defects after sintering, reducing its density and mechanical properties.

[0126] Viscosity threshold: 1000 to 5000 mPa.s. Specifically, the viscosity of the porcelain slurry reflects its fluidity and forming performance. If the viscosity is too low, the porcelain slurry is prone to delamination, and is prone to deformation during printing; if the viscosity is too high, it is not easy to extrude and form, and is prone to air bubbles. Generally, the viscosity of the porcelain slurry is controlled in the range of 1000 to 5000 mPa.s, which is more appropriate.

[0127] Solid content threshold: 55±3%. Specifically, the solid content represents the content of ceramic powder particles in the porcelain slurry. Too low solid content will reduce the density of the porcelain body, and too high solid content will easily cause the porcelain slurry to be non-uniform.

[0128] Particle size threshold: average particle size 100 nm to 300 nm, median particle size D50 not more than 200 nm, and 90% of the particles D90 not more than 300 nm. Specifically, the uniformity and fineness of the particle size are also the key to ensure the quality of the porcelain slurry. The average particle size should be between 100 nm and 300 nm to ensure sintering activity.

[0129] Binder agglomerated particle threshold: no visible agglomerated block, and no agglomerated particle more than 500 nm at any position.

[0130] Residual solid particle threshold: no more than 3 particles per square centimeter of surface, and the residual solid particle size is less than 500 nm.

[0131] Porcelain slurry impurity particle threshold: no more than 3 particles per square centimeter of surface, and the residual solid particle size is less than 500 nm.

[0132] In addition to the above parameters, the agglomerated particles, residual solid particles and impurity particles that may exist in the porcelain slurry also need to be focused on. The binder agglomerated particles are caused by uneven dispersion, which will cause uneven density and defects of the porcelain body; the residual solid particles are mainly from the wear of the grinding balls during the ball milling process, which will form different phases with different hardness in the porcelain body; the impurity particles may come from raw materials or processing environment, which will also reduce the performance of the porcelain body. Therefore, the porcelain slurry should have no visible agglomerated block, and there should be no agglomerated particles more than 500 nm at any position; the number of residual solid particles and impurity particles per square centimeter of surface should not exceed 3, and the particle size should be less than 500 nm.

[0133] Only by comprehensively evaluating the key parameters such as bubbles, viscosity, solid content, particle type and particle size, and comparing with the preset judgment standard threshold, can the quality of the porcelain slurry be objectively and comprehensively judged. At the same time, these parameters are not independent, but are interrelated and influence each other. For example, when the solid content is too high, the viscosity will rise sharply, and the bubble generation will be aggravated; and the particle agglomeration will change the solid content and the viscosity. Therefore, the judgment of the quality of the porcelain slurry needs to be made from the perspective of multiple parameters, and a systematic analysis and dynamic control are needed to obtain reliable evaluation results, guide the optimization and improvement of the production process, and finally ensure the performance and quality of the porcelain product and improve the reliability and stability of the multilayer ceramic capacitor. This is of great significance to the development of the electronic information industry.

[0134] The present application significantly improves the judgment accuracy of the quality of the porcelain slurry by optimizing the particle recognition algorithm and the multi-parameter comprehensive analysis. First, the particle mask model constructed based on deep learning can accurately segment the particles in the microscope image, and combined with morphological operation and watershed algorithm, the problems of particle adhesion and edge noise are effectively solved, and the accurate extraction of the outline of the independent particle is ensured. Second, the particle size calculation model with shape factor dynamic correction is introduced, which comprehensively considers the actual morphological characteristics of the particles, and greatly improves the reliability of the particle size data. In addition, the texture features are extracted by the gray level co-occurrence matrix, and combined with the multi-dimensional data such as particle outline and aspect ratio, the efficient classification of ceramic powder particles, agglomerated particles, residual solid particles and impurity particles is realized.

[0135] Finally, the systematic judgment standard integrates the multi-parameter data such as bubbles, viscosity, solid content, particle type and particle size, forming a comprehensive quality evaluation system, avoiding the one-sidedness of a single index. This method not only improves the automation level of detection and reduces the error of human intervention, but also quickly locates the abnormal production parameters of unqualified porcelain slurry, providing a reliable basis for process optimization, so as to ensure the high performance and stability of the multilayer ceramic capacitor.

[0136] On the basis of the above embodiment, an optional embodiment of the present application is as shown in Figure 3 The specific network information of the self-defined backbone network is as follows.

[0137] The self-defined backbone network comprises an initial convolutional layer, a first maximum pooling layer, a first dense block, a first attention mechanism, a second dense block and a second attention mechanism connected in sequence.

[0138] The initial convolutional layer comprises a first convolutional layer with a convolutional kernel number of 64, a convolutional kernel size of 7x7, a convolutional step of 2 and a convolutional padding of 3, and a first batch normalization layer and a first ReLU activation layer connected in sequence.

[0139] The pooling kernel size of the first maximum pooling layer is 3x3, the pooling step is 2, and the pooling padding is 1.

[0140] The input channel number of the first dense block is 64, the number of layers is 6, and the growth rate is 32. Each DenseLayer structure of the first dense block includes a first BatchNormalization batch normalization layer, a second ReLU activation layer, and a second convolutional layer with a convolutional kernel number of 32, a convolutional kernel size of 3x3, and a convolutional padding of 1, which are sequentially connected.

[0141] The first attention mechanism includes a first spatial attention module and a first channel attention module which are sequentially connected. The first spatial attention module includes a first spatial dimension global average pooling layer and a first spatial dimension global maximum pooling layer connected to the first dense block, a first fully connected layer connected to the first spatial dimension global average pooling layer and the first spatial dimension global maximum pooling layer, a third ReLU activation layer connected to the first fully connected layer, a second fully connected layer connected to the third ReLU activation layer, a first addition layer connected to the second fully connected layer, a first Sigmoid activation function layer connected to the first addition layer, and a first multiplication layer connected to the first dense block and the first Sigmoid activation function layer. The first channel attention module includes a first channel dimension global maximum pooling and a first channel dimension global average pooling connected to the first multiplication layer, a second addition layer connected to the first channel dimension global maximum pooling and the first channel dimension global average pooling, a third convolutional layer connected to the second addition layer, a second Sigmoid activation function layer connected to the third convolutional layer, and a second multiplication layer connected to the first multiplication layer and the second Sigmoid activation function layer.

[0142] The input channel number of the second dense block is 64+6*32, the number of layers is 12, and the growth rate is 32. Each DenseLayer structure of the second dense block includes a second BatchNormalization batch normalization layer, a fourth ReLU activation layer, and a fourth convolutional layer with a convolutional kernel number of 32, a convolutional kernel size of 3x3, and a convolutional padding of 1, which are sequentially connected.

[0143] The second attention mechanism includes a second spatial attention module and a second channel attention module connected in sequence. The second spatial attention module includes a second spatial dimension global average pooling layer and a second spatial dimension global maximum pooling layer connected to the second dense block, a third fully connected layer connected to the second spatial dimension global average pooling layer and the second spatial dimension global maximum pooling layer, a fifth ReLU activation layer connected to the third fully connected layer, a fourth fully connected layer connected to the fifth ReLU activation layer, a third addition layer connected to the fourth fully connected layer, a third Sigmoid activation function layer connected to the third addition layer, and a third multiplication layer connected to the second dense block and the first Sigmoid activation function layer. The second channel attention module includes a second channel dimension global maximum pooling and a second channel dimension global average pooling connected to the third multiplication layer, a fourth addition layer connected to the second channel dimension global maximum pooling and the second channel dimension global average pooling, a fifth convolutional layer connected to the fourth addition layer, a fourth Sigmoid activation function layer connected to the fifth convolutional layer, and a fourth multiplication layer connected to the third multiplication layer and the fourth Sigmoid activation function layer.

[0144] The design of the customized backbone network embodies the efficiency and flexibility of the convolutional neural network in feature extraction. The initial convolutional layer can effectively capture low-level features of the image. The first maximum pooling layer further compresses the feature map size, reduces the calculation amount, and preserves important features. The first dense block enhances the reusability of features through dense connection, so that the network can retain and transmit low-level features in a deeper level, thereby improving the efficiency of feature extraction. The first attention mechanism is composed of a first spatial attention module and a first channel attention module. The first spatial attention module extracts spatial features first. The first channel attention module further enhances the feature selection capability. The second dense block can further expand the diversity and complexity of the features in a deeper layer, and improve the expression ability of the network. The second attention mechanism is composed of a second spatial attention module and a second channel attention module. The second spatial attention module extracts spatial features again. The second channel attention module further optimizes feature selection.

[0145] The design of this customized backbone network realizes multi-level feature extraction and optimization through the combination of dense blocks and attention mechanisms, which can effectively improve the performance of the network in complex tasks. In the target detection task, the network can more accurately capture the spatial and channel features of the target, thereby improving the detection accuracy.

[0146] On the basis of the above-mentioned embodiments, in an optional embodiment of the present application, as Figure 4As shown, the region proposal network comprises a first shared convolutional layer connected to the customized backbone network, a sixth ReLU activation layer connected to the first shared convolutional layer, a sixth convolutional layer and a seventh convolutional layer respectively connected to the sixth ReLU activation layer, a first Softmax function layer connected to the sixth convolutional layer, and a non-maximum suppression layer connected to the first Softmax function layer and the seventh convolutional layer.

[0147] The region proposal network generates target candidate regions from the features extracted by the customized backbone network. The first shared convolutional layer first performs convolution operation on the input features to extract spatial information. The sixth ReLU activation layer introduces nonlinearity to enhance the expression ability of the features. The seventh convolutional layer and the sixth convolutional layer are respectively responsible for generating position information and class confidence of the candidate regions. The sixth convolutional layer outputs the class score of each candidate region. The seventh convolutional layer outputs four coordinate values of each candidate region, representing the position of the bounding box. The first Softmax function layer converts the class score into a probability distribution, indicating the probability of each candidate region belonging to each target class. The non-maximum suppression layer removes redundant candidate regions according to the class score and position information, and retains the most representative proposals.

[0148] Preferably, as shown in the figure, Figure 5 As shown, the detection branch comprises a first RoI pooling layer connected to the customized backbone network and the region proposal network, a fifth fully connected layer connected to the first RoI pooling layer, a seventh ReLU activation layer connected to the fifth fully connected layer, a sixth fully connected layer connected to the seventh ReLU activation layer, an eighth ReLU activation layer connected to the sixth fully connected layer, and a bounding box regressor connected to the eighth ReLU activation layer. The first RoI pooling layer is adapted to output a feature map with a size of 7x7, which is used to align the candidate regions output by the region proposal network and the extracted features output by the customized backbone network. The input dimension of the fifth fully connected layer is the size of the RoI pooling layer output, and the output dimension is 1024. The input dimension of the sixth fully connected layer is 1024, and the output dimension is 1024. The input dimension of the bounding box regressor is 1024, and the output dimension is 4 to represent the diagonal coordinates of the bounding box.

[0149] The detection branch performs target detection through the candidate regions generated by the region proposal network. The first RoI pooling layer unifies candidate regions of different sizes into a fixed-size feature map, facilitating subsequent processing. The fifth fully connected layer flattens the 7x7 feature map into a high-dimensional vector for feature fusion. The seventh ReLU activation layer introduces nonlinearity to enhance the feature expression ability. The sixth fully connected layer further extracts features to enhance the discriminative ability of the model. The eighth ReLU activation layer again introduces nonlinearity to improve the quality of the features. The bounding box regressor outputs the refined coordinates of the bounding box according to the extracted features to improve the detection accuracy.

[0150] Preferred, such as Figure 5 As shown, the mask branch includes the following layers connected in sequence: a second RoI pooling layer with an output feature map size of 14x14; an eighth convolutional layer with 256 kernels, a kernel size of 3x3, and a padding of 1; a ninth ReLU activation layer with 256 kernels, a kernel size of 3x3, and a padding of 1; a tenth ReLU activation layer with 256 kernels, a kernel size of 3x3, and a padding of 1; an eleventh ReLU activation layer with 256 kernels, a kernel size of 3x3, and a padding of 1; a twelfth ReLU activation layer with 256 kernels, a kernel size of 3x3, and a padding of 1; a first deconvolutional layer with 256 kernels, a kernel size of 2x2, and a stride of 2; a thirteenth ReLU activation layer; a first final convolutional layer with 1 kernel and a kernel size of 1x1; and a fifth Sigmoid activation function layer. The second RoI pooling layer is connected to the custom backbone network and the detection branch.

[0151] The mask branch segments the target by detecting the bounding boxes generated by the branch. The second RoI pooling layer unifies the features within the bounding boxes into a fixed-size feature map, facilitating segmentation. The eighth convolutional layer extracts spatial features, enhancing feature representation. The ninth ReLU activation layer introduces non-linearity, improving feature quality. The ninth convolutional layer further extracts features, enhancing the model's discriminative ability. The tenth ReLU activation layer again introduces non-linearity, improving feature quality. The tenth convolutional layer continues to extract features, enhancing the model's expressive ability. The eleventh ReLU activation layer introduces non-linearity, improving feature quality. The eleventh convolutional layer extracts high-level features, enhancing the model's discriminative ability. The twelfth ReLU activation layer introduces non-linearity, improving feature quality. The first deconvolutional layer upsamples the feature map, restoring spatial resolution. The thirteenth ReLU activation layer introduces non-linearity, improving feature quality. The first final convolutional layer outputs a binary mask of the target, depicting the target contour. The fifth sigmoid activation function layer converts the output value into a probability between 0 and 1, representing the probability that each pixel belongs to the target.

[0152] Preferably, the first output layer is adapted to output bounding box information and mask information. The bounding box information is a 4-dimensional vector for each detected particle, representing the coordinates of the top-left and bottom-right corners of the bounding box. The mask information is a binary mask proportional to the size of the input image for each detected particle, used to delineate the particle's outline.

[0153] The first output layer outputs the bounding box information and the mask information, and provides information for subsequent detection and segmentation tasks. The bounding box information represents the position of the target through a four-dimensional vector, facilitating target positioning. The mask information represents the contour of the target through a binary mask, facilitating target segmentation. The output of this information provides accurate target information for subsequent tasks, improving the accuracy of detection and segmentation.

[0154] On the basis of the above-mentioned embodiments, in an optional embodiment of the present application, as shown in Figure 6 The particle classification model includes an image feature processing network for processing particle contours, a numerical feature processing network for processing aspect ratio, area, perimeter, shape factor, particle size and texture features, and a feature fusion classification network connected to the image feature processing network and the numerical feature processing network.

[0155] The image feature processing network includes a twelfth convolutional layer, a ninth ReLU activation layer, a second max-pooling layer, a thirteenth convolutional layer, a tenth ReLU activation layer, a third max-pooling layer, a fourteenth convolutional layer, an eleventh ReLU activation layer, a fourth max-pooling layer, and a Flatten flattening layer connected in sequence. Among them, the Flatten flattening layer is adapted to convert the two-dimensional feature map after three times of convolution and pooling into a one-dimensional vector.

[0156] The image feature processing network of the particle classification model extracts features from the particle contour image through a multi-layer convolutional neural network. First, the twelfth convolutional layer is used to perform convolution operation on the input particle contour image to extract preliminary image features. Then, the feature map is subjected to nonlinear transformation through the ninth ReLU activation layer, and down-sampling is performed through the second max-pooling layer to reduce the feature map size. Then, the thirteenth convolutional layer is used to further extract image features, and then the ReLU activation and max-pooling are performed again. Finally, the fourteenth convolutional layer is used to extract high-level semantic features, and the ReLU activation and max-pooling are performed. After three times of convolution and pooling, a feature map with smaller size but rich semantic information is obtained. The Flatten flattening layer converts the two-dimensional feature map into a one-dimensional vector, facilitating subsequent processing of the fully connected network.

[0157] Suppose the feature map output by the fourth max-pooling layer has a size of w x h x c. Wherein, w is the width, h is the height, and c is the number of channels. The Flatten layer converts it into a one-dimensional vector of 1 x (w x h x c). For example, if the feature map size is 7 x 7 x 512, it becomes a vector of 1 x 25088 after flattening. This transformation not only reduces the data dimension, but also preserves the spatial information of the original feature map.

[0158] The numerical feature processing network includes a seventh fully connected layer and a twelfth ReLU activation layer connected in sequence. Among them, the seventh fully connected layer is adapted to connect multiple numerical features.

[0159] The numerical feature processing network is used to process numerical features of the particle, such as aspect ratio, area, perimeter, etc. The seventh fully connected layer connects these numerical features into a high-dimensional vector as a comprehensive representation of the numerical features. For example, the particle has 6 numerical features, each of which is normalized to range between 0 and 1, and the input of the fully connected layer is a 1x6 vector. If the fully connected layer has 1024 neurons, the weight matrix size is 6x1024, the bias vector is 1x1024, and the output feature vector is 1x1024. The twelfth ReLU activation layer performs a nonlinear transformation on the output of the fully connected layer to improve the feature representation capability. The ReLU function sets the value less than 0 to 0 and keeps the value greater than 0 unchanged. It is simple to calculate and can alleviate the gradient disappearance problem and speed up the network convergence.

[0160] The feature fusion classification network includes a concatenation layer, an eighth fully connected layer, a thirteenth ReLU activation layer, a ninth fully connected layer, a fourteenth ReLU activation layer, and a second Softmax function layer connected in sequence. The concatenation layer is adapted to concatenate the outputs of the image feature processing network and the numerical feature processing network. The second Softmax function layer outputs the probabilities of the particle belonging to four particle types.

[0161] The feature fusion classification network fuses the image features and the numerical features and predicts the particle category. The concatenation layer concatenates the image feature vector and the numerical feature vector in the column dimension. For example, the image feature vector is 1x25088 and the numerical feature vector is 1x1024, and the concatenated vector is 1x26112. The eighth fully connected layer and the thirteenth ReLU activation layer further extract the fusion features. The ninth fully connected layer reduces the feature dimension to the number of categories, such as reducing the feature dimension from 512 to 4 in a four-classification task. The fourteenth ReLU activation layer increases the nonlinear representation capability. The last layer of the second Softmax function converts the output of the fully connected layer into a probability distribution, indicating the probabilities of the sample belonging to each particle type under the current input. The greater the probability value, the greater the possibility of the sample belonging to the category. The Softmax function converts the output into a meaningful probability distribution, which is convenient for interpreting the result.

[0162] On the basis of the above-mentioned embodiment, in an optional embodiment of the present application, the porcelain slurry production monitoring method further includes steps S08 to S10.

[0163] S08, when it is judged that the porcelain slurry is unqualified, obtaining the related parameters of the porcelain slurry. The related parameters include production parameters that can be controlled in the production process and characteristic parameters of the porcelain slurry after production. The production parameters include production personnel information, production machine information, production material information, raw material ratio information, stirring speed, stirring time, ingredient room temperature, ingredient room humidity, porcelain slurry container temperature, and porcelain slurry container humidity. The characteristic parameters include bubble information, viscosity, solid content, and the particle information.

[0164] S09, inputting the related parameters of the unqualified porcelain slurry into an unqualified reason analysis model to obtain an unqualified production parameter set that causes the porcelain slurry to be unqualified. Preferably, step S09 specifically includes steps S091 to S092.

[0165] S61, obtaining the maximum value and the minimum value of each related parameter, and performing normalization processing on each related parameter of the unqualified porcelain slurry according to the maximum value and the minimum value to obtain fault input data of the unqualified reason analysis model.

[0166] S62, inputting the fault input data into the unqualified reason analysis model to obtain production parameters that cause the porcelain slurry to be unqualified, which are defined as an unqualified production parameter set.

[0167] The production parameters that cause the porcelain slurry to be unqualified are accurately judged through the unqualified reason analysis model, which significantly improves the accuracy of problem diagnosis. This method can quickly identify problem parameters, thereby greatly reducing downtime analysis time and improving the operation efficiency of the production line. By accurately and quickly finding the source of the problem and providing a solution, the scheme effectively reduces economic losses caused by downtime.

[0168] On the basis of the above embodiment, in an optional embodiment of the present application, as shown in Figure 7 The unqualified reason analysis model is an M-tag independent classification model. The network structure of the unqualified reason analysis model includes:

[0169] A second input layer is provided with M neurons. M is the number of related parameters, and each related parameter corresponds to a neuron.

[0170] A first hidden layer is provided with 4M neurons, and the first hidden layer uses a ReLU activation function.

[0171] A first random inactivation layer, and the inactivation rate of the first random inactivation layer is 0.2.

[0172] A second hidden layer is provided with 4M neurons, and the second hidden layer uses a ReLU activation function.

[0173] A second random inactivation layer, and the inactivation rate of the second random inactivation layer is 0.2.

[0174] The second output layer is provided with M neurons and M linear activation functions, respectively, to output M unnormalized prediction values of the related parameters.

[0175] The conditional random field is provided with an MxM transition matrix to represent the transition probability between any two labels.

[0176] The first and second random inactivation layers are deleted after the model is trained.

[0177] The existing multi-label independent classification model has the following disadvantages in the existing multi-layer perceptron: 1. Without a large enough data set or a proper regularization method, it is easy to fall into overfitting, which leads to good performance on the training set but poor generalization ability on the test set. 2. When processing high-dimensional input data, the computational complexity is high, and it is difficult to effectively learn the potential features in the data. 3. Label correlation is difficult to capture: in the multi-label classification task, there may be correlation or dependence between labels. The basic independent classification model is difficult to directly capture the complex relationship between these labels and cannot make accurate judgments.

[0178] Therefore, the embodiment of the application adds a random inactivation layer between the hidden layer and the output layer, thereby effectively reducing overfitting.

[0179] In addition, in general, the abnormality of the characteristic parameters of the product is not caused by one parameter, but by the joint action of several production parameters. Therefore, in the embodiment, a special CRF layer (i.e., a conditional random field layer) is added after the output layer, which is called "label pair CRF". This CRF layer will construct an MxM transition matrix to represent the transition probability between any two labels. The elements in the transition matrix represent the probability or preference of a label transitioning to another label. For each sample, the CRF layer will calculate the most likely label combination according to the emission score output by the independent classification model and the transition matrix. Thus, the model considers the mutual influence between all labels. Further, it is more accurate to determine which production parameters jointly cause the product defect. Specifically, each neuron of the output layer is responsible for calculating the emission score (logits) of the corresponding label, and then inputting the linear activation function to the conditional random field layer. The emission score reflects the possibility of each label given the input.

[0180] ​S10, input the related parameters and the poor production parameters into a poor cause solving model to obtain a recommended value of the poor production parameters. Specifically, the related parameters also need to be normalized before being input into the fault solving model.

[0181] The method for monitoring the production of porcelain slurry of a multilayer ceramic capacitor according to the application can accurately determine the process parameters that cause poor capacitance through the poor cause solving model, thereby significantly improving the accuracy of problem diagnosis. This method can quickly identify problem parameters, thereby greatly reducing downtime analysis time and improving the operation efficiency of the production line. By accurately and quickly finding the source of the problem and providing a solution, the scheme effectively reduces economic losses caused by downtime.

[0182] Based on the above-mentioned embodiments, in an optional embodiment of the application, as shown in Figure 8 The network structure of the poor cause solving model comprises:

[0183] The third input layer is provided with M neurons corresponding to the M related parameters, and uses a linear activation function.

[0184] The mask input layer is provided with M neurons for receiving a mask vector corresponding to the input data. Each element is 0 or 1, 0 indicating that the corresponding related parameter is marked as abnormal, and 1 indicating normal.

[0185] The encoder layer comprises a first mask application layer, a first hidden layer, a second hidden layer and a bottleneck layer connected in sequence. The first mask application layer is used to multiply the data of the input layer with the mask vector, so as to ensure that only the parameters marked as abnormal are set to zero, and the normal parameters remain unchanged. The first hidden layer is provided with 128 neurons and uses a ReLU activation function. The second hidden layer is provided with 64 neurons and uses a ReLU activation function. The bottleneck layer is provided with 32 neurons and uses a ReLU activation function.

[0186] The decoder layer comprises a third hidden layer and a fourth hidden layer. The third hidden layer is provided with 64 neurons and uses a LeakyReLU activation function. The fourth hidden layer is provided with 128 neurons and uses a LeakyReLU activation function.

[0187] The second mask application layer is used to apply the mask vector, so as to ensure that only the parameters marked as abnormal are updated by the model, and the other parameters remain unchanged.

[0188] The third output layer is provided with M neurons for reconstructing the input data, and uses a linear activation function to output the recommended value of the related parameters in the original scale.

[0189] Specifically, the encoder compresses the input data (relevant parameters) into a low-dimensional representation, called the latent representation. The decoder then reconstructs the original data from this latent representation. During training, the autoencoder learns how to reconstruct the input data with minimal error. When an abnormal production parameter occurs, the dataset containing the abnormal parameter is input into the trained autoencoder, which attempts to reconstruct the abnormal parameter from the representations of other normal parameters in the latent space, thereby obtaining an estimate of the normal value of the abnormal parameter.

Claims

1. A method of monitoring production of a ceramic slurry for a multilayer ceramic capacitor, characterized by, The method comprises the following steps: obtaining a porcelain slurry sample; using an ultrasonic detector to detect bubble information of the porcelain slurry sample; using a rotary viscometer to measure the viscosity of the porcelain slurry sample at a preset temperature and a preset rotating speed; taking a porcelain slurry sample of a preset mass to dry, and calculating the mass percentage of the solid residue after drying to obtain the solid content of the porcelain slurry sample; preparing the porcelain slurry sample into a dry smear, and obtaining a microscope image of the smear; identifying particle information in the microscope image through a particle identification model, wherein the particle information comprises particle types and particle sizes of each particle; referring to a preset judgment standard, judging whether the porcelain slurry is qualified according to the bubble information, the viscosity, the solid content, the particle types and the particle sizes.

2. The method for monitoring the production of ceramic slurry for a multilayer ceramic capacitor according to claim 1, wherein The particle information in the microscope image is identified through the particle identification model, specifically comprising: preprocessing the microscope image; obtaining mask information of each particle in the preprocessed microscope image through a particle mask model; performing particle separation and morphological operation according to the mask information to obtain independent particle contours; obtaining particle sizes of each particle according to the independent particle contours; extracting multiple features of each particle and classifying them according to the mask information, the particle sizes and the preprocessed microscope image to obtain particle types of each particle.

3. The method of claim 2, wherein the slurry is a ceramic slurry for a multilayer ceramic capacitor. The particle mask model comprises a first input layer, a self-defined backbone network connected to the first input layer, a region proposal network connected to the self-defined backbone network, a detection branch connected to the self-defined backbone network and the region proposal network, a mask branch connected to the self-defined backbone network and the detection branch, and a first output layer connected to the mask branch. The mask information of each particle in the preprocessed microscope image is obtained through the particle mask model, specifically comprising: inputting the preprocessed microscope image into the self-defined backbone network through the first input layer to extract a feature map; inputting the feature map extracted by the self-defined backbone network into the region proposal network to generate a candidate region possibly containing a particle on the feature map; judging whether the candidate region contains a particle through the detection branch according to the candidate region; generating a binary mask for the candidate region containing a particle using the mask branch to segment and obtain the specific contour of the particle in the image, and obtaining the mask information of each particle; outputting the mask information and position information of the particle through the first output layer.

4. The method of claim 3, wherein the slurry is a ceramic slurry for a multilayer ceramic capacitor. The self-defined backbone network comprises an initial convolution layer, a first max-pooling layer, a first dense block, a first attention mechanism, a second dense block and a second attention mechanism connected in sequence; the initial convolution layer comprises a first convolution layer with a convolution kernel number of 64, a convolution kernel size of 7x7, a convolution step of 2 and a convolution padding of 3, and a first batch normalization layer and a first ReLU activation layer connected in sequence; the first max-pooling layer has a pooling kernel size of 3x3, a pooling step of 2 and a pooling padding of 1; The first dense block has an input channel number of 64, a layer number of 6, and a growth rate of 32; each DenseLayer structure of the first dense block includes a first BatchNormalization batch normalization layer, a second ReLU activation layer, and a second convolutional layer with a convolutional kernel number of 32, a convolutional kernel size of 3x3, and a convolutional padding of 1, which are sequentially connected; The first attention mechanism includes a first spatial attention module and a first channel attention module which are sequentially connected; The first spatial attention module includes a first spatial dimension global average pooling layer and a first spatial dimension global maximum pooling layer connected to the first dense block, a first fully connected layer connected to the first spatial dimension global average pooling layer and the first spatial dimension global maximum pooling layer, a third ReLU activation layer connected to the first fully connected layer, a second fully connected layer connected to the third ReLU activation layer, a first addition layer connected to the second fully connected layer, a first Sigmoid activation function layer connected to the first addition layer, and a first multiplication layer connected to the first dense block and the first Sigmoid activation function layer; the first channel attention module includes a first channel dimension global maximum pooling and a first channel dimension global average pooling connected to the first multiplication layer, a second addition layer connected to the first channel dimension global maximum pooling and the first channel dimension global average pooling, a third convolutional layer connected to the second addition layer, a second Sigmoid activation function layer connected to the third convolutional layer, and a second multiplication layer connected to the first multiplication layer and the second Sigmoid activation function layer; The second dense block has an input channel number of 64+6*32, a layer number of 12, and a growth rate of 32; each DenseLayer structure of the second dense block includes a second BatchNormalization batch normalization layer, a fourth ReLU activation layer, and a fourth convolutional layer with a convolutional kernel number of 32, a convolutional kernel size of 3x3, and a convolutional padding of 1, which are sequentially connected; The second attention mechanism includes a second spatial attention module and a second channel attention module which are sequentially connected; The second spatial attention module comprises a second spatial dimension global average pooling layer and a second spatial dimension global maximum pooling layer connected to the second dense block, a third fully connected layer connected to the second spatial dimension global average pooling layer and the second spatial dimension global maximum pooling layer, a fifth ReLU activation layer connected to the third fully connected layer, a fourth fully connected layer connected to the fifth ReLU activation layer, a third addition layer connected to the fourth fully connected layer, a third Sigmoid activation function layer connected to the third addition layer, and a third multiplication layer connected to the second dense block and the first Sigmoid activation function layer; the second channel attention module comprises a second channel dimension global maximum pooling and a second channel dimension global average pooling connected to the third multiplication layer, a fourth addition layer connected to the second channel dimension global maximum pooling and the second channel dimension global average pooling, a fifth convolutional layer connected to the fourth addition layer, a fourth Sigmoid activation function layer connected to the fifth convolutional layer, and a fourth multiplication layer connected to the third multiplication layer and the fourth Sigmoid activation function layer.

5. The method of claim 3, wherein the slurry is a ceramic slurry for a multilayer ceramic capacitor. The region proposal network comprises a first shared convolutional layer connected to the customized backbone network, a sixth ReLU activation layer connected to the first shared convolutional layer, a sixth convolutional layer and a seventh convolutional layer connected to the sixth ReLU activation layer respectively, a first Softmax function layer connected to the sixth convolutional layer, and a non-maximum suppression layer connected to the first Softmax function layer and the seventh convolutional layer. The detection branch comprises a first RoI pooling layer connected to the customized backbone network and the region proposal network, a fifth fully connected layer connected to the first RoI pooling layer, a seventh ReLU activation layer connected to the fifth fully connected layer, a sixth fully connected layer connected to the seventh ReLU activation layer, an eighth ReLU activation layer connected to the sixth fully connected layer, and a bounding box regressor connected to the eighth ReLU activation layer; the first RoI pooling layer is adapted to output a feature map with a size of 7x7, so as to align the candidate regions output by the region proposal network and the extracted features output by the customized backbone network; The input dimension of the fifth fully connected layer is the size of the RoI pooling layer output, and the output dimension is 1024; the input dimension of the sixth fully connected layer is 1024, and the output dimension is 1024; the input dimension of the bounding box regressor is 1024, and the output dimension is 4 to represent the diagonal coordinates of the bounding box; The mask branch comprises, connected in sequence: a second RoI pooling layer with an output feature map size of 14x14, an eighth convolutional layer with a convolutional kernel number of 256, a convolutional kernel size of 3x3, and a convolutional padding of 1, a ninth ReLU activation layer, a ninth convolutional layer with a convolutional kernel number of 256, a convolutional kernel size of 3x3, and a convolutional padding of 1, a tenth ReLU activation layer, a tenth convolutional layer with a convolutional kernel number of 256, a convolutional kernel size of 3x3, and a convolutional padding of 1, an eleventh ReLU activation layer, an eleventh convolutional layer with a convolutional kernel number of 256, a convolutional kernel size of 3x3, and a convolutional padding of 1, a twelfth ReLU activation layer, a first deconvolutional layer with a convolutional kernel number of 256, a convolutional kernel size of 2x2, and a convolutional step of 2, a thirteenth ReLU activation layer, a first final convolutional layer with a convolutional kernel number of 1 and a convolutional kernel size of 1x1, and a fifth Sigmoid activation function layer; the second RoI pooling layer is connected to the self-defined backbone network and the detection branch; The first output layer is adapted to output bounding box information and mask information; the bounding box information is that, for each detected particle, a 4-dimensional vector is output to represent the coordinates of the upper left corner and the lower right corner of the bounding box; the mask information is that, for each detected particle, a binary mask proportional to the size of the input image is output to depict the outline of the particle.

6. The method of claim 2, wherein the slurry is a ceramic slurry for a multilayer ceramic capacitor. According to the mask information, particle separation and morphological operation are performed to obtain independent particle outlines, specifically including: According to the mask information, morphological opening operation is performed to first detect whether there is a hollow region inside the particle, and if there is a hollow, morphological closing operation is used for filling, then a corrosion operation is used to remove noise on the edge of the mask, and then an inflation operation is used to restore the shape of the particle, and finally particles with a pixel number less than a first preset pixel threshold are deleted to obtain morphologically processed mask information; Distance transformation is performed on the morphologically processed mask information to calculate the Euclidean distance of each foreground pixel to the nearest background pixel to obtain a gray-scale image after distance transformation; wherein the value of each pixel of the gray-scale image represents the distance to the background; According to the gray-scale image, a local maximum value detection algorithm is used to find the center point of each particle, then a blank image with the same size as the input image is created, and the center point is marked in the blank image to generate a particle marker image; wherein each center point in the particle marker image represents an independent particle; According to the gray-scale image after distance transformation and the particle marker image, a watershed algorithm is used to segment the gray-scale image to separate the particles adhered in the mask image to obtain the outline of each single particle; Whether the outlines of two adjacent single particles are connected is detected; For the contours of the two connected individual particles, the edge smoothness of the two independent contours and the edge smoothness of the merged contour of the outer contour after the two contours are merged are calculated respectively; when the edge smoothness of the merged contour is less than the edge smoothness of the two independent contours, the two connected contours are merged; finally, the contours with a number of pixels less than a first preset pixel threshold are deleted, and the independent particle contour is obtained; The calculation model of the edge smoothness is: ； ； ； ； wherein is the edge smoothness, is the number of points on the contour, is the number of points on the contour, is the change in curvature between adjacent points, is the curvature of the point, is the curvature of the point, is the vector angle at the point, is the distance between the point and the point, is the distance between the point and the point, is the coordinate of the point, is the coordinate of the point, is the coordinate of the point, is the vector from the point to the point, is the vector from the point to the point.

7. The method of claim 2 to 6, wherein According to the independent particle contour, the particle size of each particle is obtained, specifically including: According to the independent particle contour, the major axis length and the minor axis length are determined by the minimum circumscribed rectangle algorithm, and the aspect ratio of the particle is calculated; According to the magnification of the microscope image and the independent particle contour, the area and the perimeter of the contour of each particle are obtained; wherein the area is calculated based on the pixel counting method; the perimeter is calculated by the pixel spacing method; According to the area and the perimeter, the particle size is obtained under the consideration of the shape factor; wherein the calculation model of the particle size is: ； ； wherein is the particle size of the particles, is the area, is the circumference, is the shape factor.

8. The method of claim 7, wherein the slurry is a ceramic slurry for a multilayer ceramic capacitor. According to the mask information, the particle size and the preprocessed microscope image, a plurality of features of each particle are extracted and classified respectively, and the particle type of each particle is obtained, specifically including: According to the mask information and the preprocessed microscope image, the texture features are extracted by the gray level co-occurrence matrix; wherein the texture features include contrast, correlation and energy; The particle contour, the aspect ratio, the area, the perimeter, the shape factor, the particle size of the particle, and the texture features are input into the particle classification model to obtain the particle type; wherein the particle type includes ceramic powder particles, porcelain slurry agglomerated particles, residual solid particles, and porcelain slurry impurity particles; The particle classification model includes an image feature processing network for processing the particle contour, a numerical feature processing network for processing the aspect ratio, the area, the perimeter, the shape factor, the particle size and the texture features, and a feature fusion classification network connected to the image feature processing network and the numerical feature processing network; The image feature processing network includes a twelfth convolutional layer, a ninth ReLU activation layer, a second maximum pooling layer, a thirteenth convolutional layer, a tenth ReLU activation layer, a third maximum pooling layer, a fourteenth convolutional layer, an eleventh ReLU activation layer, a fourth maximum pooling layer, and a Flatten flattening layer connected in turn; wherein the Flatten flattening layer is adapted to convert the two-dimensional feature map after three times of convolution and pooling into a one-dimensional vector; The numerical feature processing network includes a seventh fully connected layer and a twelfth ReLU activation layer connected in turn; wherein the seventh fully connected layer is adapted to connect a plurality of numerical features. The feature fusion classification network comprises a concatenation layer, an eighth fully connected layer, a thirteenth ReLU activation layer, a ninth fully connected layer, a fourteenth ReLU activation layer and a second Softmax function layer connected in sequence; wherein the concatenation layer is adapted to concatenate the outputs of the image feature processing network and the numerical feature processing network together; and the second Softmax function layer outputs the probabilities of the particles belonging to the four particle types.

9. The method of claim 1 to 6, wherein The particle types include ceramic powder particles, porcelain slurry agglomerate particles, residual solid particles, and porcelain slurry impurity particles. The preset judgment criteria include: a porcelain slurry bubble threshold value: no more than 10 bubbles per cubic centimeter, and a maximum bubble diameter of no more than 5 μm; a viscosity threshold value: 1000 to 5000 mPa.s; a solid content threshold value: 55 ± 3%; a particle size threshold value: an average particle size of 100 nm to 300 nm, a median particle size D50 of no more than 200 nm, and 90% of the particles D90 of no more than 300 nm; an adhesive agglomerate particle threshold value: no visible agglomerate blocks, and no agglomerate particles of more than 500 nm at any position; a residual solid particle threshold value: no more than 3 particles per square centimeter of surface, and a residual solid particle size of less than 500 nm; a porcelain slurry impurity particle threshold value: no more than 3 particles per square centimeter of surface, and a residual solid particle size of less than 500 nm.

10. The method of claim 7, wherein the slurry is a ceramic slurry for a multilayer ceramic capacitor. The porcelain slurry production monitoring method further includes: when it is determined that the porcelain slurry is unqualified, obtaining relevant parameters of the porcelain slurry; wherein the relevant parameters include production parameters that can be controlled in the production process, and characteristic parameters of the porcelain slurry itself after production; the production parameters include production personnel information, production machine information, production material information, raw material ratio information, stirring speed, stirring time, ingredient room temperature, ingredient room humidity, porcelain slurry container temperature, and porcelain slurry container humidity; and the characteristic parameters include bubble information, viscosity, solid content, and the particle information; inputting the relevant parameters of the unqualified porcelain slurry into a defect cause analysis model to obtain a set of unqualified production parameters; inputting the relevant parameters and the set of unqualified production parameters into a defect cause solving model to obtain suggested values of the unqualified production parameters. The defect cause analysis model is an M-tag independent classification model; and the network structure of the defect cause analysis model includes: a second input layer provided with M neurons; wherein M is the number of relevant parameters, and each relevant parameter corresponds to one neuron; a first hidden layer provided with 4M neurons, and the first hidden layer uses a ReLU activation function; a first random inactivation layer, and the inactivation rate of the first random inactivation layer is 0.2; a second hidden layer provided with 4M neurons, and the second hidden layer uses a ReLU activation function; a second random inactivation layer, and the inactivation rate of the second random inactivation layer is 0.2; a second output layer provided with M neurons and M linear activation functions, and respectively used to output unnormalized predicted values of the M relevant parameters; wherein each neuron corresponds to one process parameter. The conditional random field is provided with an MxM transition matrix for representing transition probability between any two labels; wherein the conditional random field uses a Viterbi algorithm for decoding; The first random inactivation layer and the second random inactivation layer are deleted after the model is trained; a full connection structure is used between the input layer, the first hidden layer, the second hidden layer and the output layer, and each neuron in each layer is connected with all the neurons of the previous layer; The network structure of the adverse reason solving model comprises: The third input layer is provided with M neurons corresponding to M related parameters, and uses a linear activation function; The mask input layer is provided with M neurons for receiving a mask vector corresponding to the input data; wherein each element is 0 or 1, 0 indicating that the corresponding related parameter is marked as abnormal, and 1 indicating normal; The encoder layer comprises a first mask application layer, a first hidden layer, a second hidden layer and a bottleneck layer connected in sequence; wherein the first mask application layer is used to multiply the data of the input layer with the mask vector, ensuring that only the parameters marked as abnormal are set to zero, and the normal parameters remain unchanged; the first hidden layer is provided with 128 neurons and uses a ReLU activation function; the second hidden layer is provided with 64 neurons and uses a ReLU activation function; and the bottleneck layer is provided with 32 neurons and uses a ReLU activation function; The decoder layer comprises a third hidden layer and a fourth hidden layer; wherein the third hidden layer is provided with 64 neurons and uses a LeakyReLU activation function; and the fourth hidden layer is provided with 128 neurons and uses a LeakyReLU activation function; The second mask application layer is used to apply the mask vector, ensuring that only the parameters marked as abnormal are updated by the model, and other parameters remain unchanged; The third output layer is provided with M neurons for reconstructing the input data, and uses a linear activation function to output the recommended values of the related parameters maintaining the original scale.

Images