Pressing parameter optimization method and system based on online defect detection of silicon carbide biscuit

Abstract

The invention discloses a pressing parameter optimization method and system based on silicon carbide biscuit defect online detection, and relates to the technical field of silicon carbide biscuit defect detection.The method comprises the steps that tomography scanning is conducted on a silicon carbide biscuit to be sintered, and a biscuit three-dimensional density distribution diagram is obtained; carrying out surface topography measurement to obtain a biscuit high-precision three-dimensional geometric model; carrying out geometric distortion correction to obtain a biscuit fusion defect space model; executing fine-grained defect detection to obtain a plurality of defect space distribution characteristics; outputting a plurality of single-dimensional suppression correction parameters; performing spatial multivariable coupling processing, and outputting global suppression coupling parameters; and carrying out batch pressing on the silicon carbide biscuit to be sintered. The technical problems that in the prior art, silicon carbide biscuit defect detection precision is low, pressing parameter optimization lacks pertinence, and quality fluctuation of batched pressed biscuits is large are solved, and the technical effects that silicon carbide biscuit defect online detection and pressing parameter optimization are achieved, and the pressing quality and consistency are improved are achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of silicon carbide green body defect detection, and particularly relates to a pressing parameter optimization method and system based on silicon carbide green body defect online detection. BACKGROUND

[0002] In the prior art, green body defect detection is mostly performed by a single method and a rough process, and it is difficult to simultaneously capture hidden defects such as internal density unevenness and porosity and external geometric distortion, and it is also difficult to realize accurate positioning and quantitative analysis of defects. At the same time, the correction of the pressing process has long relied on the experience of operators. When defects occur in the green body, the parameters such as pressure and pressure holding time are adjusted according to subjective experience, and there is a lack of systematic research on the correlation between defect distribution characteristics and pressing parameters. This detection and correction mode makes the defect correction lack of pertinence, and it is difficult to effectively eliminate existing defects, and new defects may be generated due to insufficient parameter adaptability, ultimately resulting in large quality fluctuations and poor consistency of the green body in batch production, and it is difficult to meet the stringent requirements of the field of semiconductor equipment and the like for high-purity and high-density silicon carbide parts.

[0003] In the prior art, the defect detection precision of the silicon carbide green body is low, the pressing parameter optimization lacks pertinence, and the quality of the batch-pressed green body fluctuates greatly. SUMMARY

[0004] The present application provides a pressing parameter optimization method and system based on silicon carbide green body defect online detection, which is used to solve the technical problems of low defect detection precision, lack of pertinence in pressing parameter optimization, and large quality fluctuations of batch-pressed green bodies in the prior art.

[0005] In view of the above problems, the present application provides a pressing parameter optimization method and system based on silicon carbide green body defect online detection.

[0006] In a first aspect of the present application, a pressing parameter optimization method based on silicon carbide green body defect online detection is provided, and the method comprises: The three-dimensional density distribution map of the green body is obtained by tomography of the green body to be sintered by an industrial CT device; the high-precision three-dimensional geometric model of the green body is obtained by surface topography measurement of the green body to be sintered by a 3D optical scanning device; the green body fusion defect space model is obtained by performing geometric distortion correction after spatial alignment and fusion of the three-dimensional density distribution map of the green body and the high-precision three-dimensional geometric model of the green body; the plurality of defect space distribution characteristics corresponding to a plurality of defect types is obtained by performing fine-grained defect detection on the green body fusion defect space model; the plurality of single-dimensional pressing correction parameters is output by performing pressing associated defect inversion according to the plurality of defect space distribution characteristics; the global pressing coupling parameter is output by performing spatial multivariate coupling processing of the plurality of single-dimensional pressing correction parameters based on the spatial attributes of the green body; and the batch pressing of the green body to be sintered is performed by using the global pressing coupling parameter.

[0007] In a second aspect of the present application, a pressing parameter optimization system based on online detection of defects of a silicon carbide green body is provided, and the system comprises: A distribution map acquisition module is configured to obtain a three-dimensional density distribution map of a green body to be sintered by tomography of the green body by an industrial CT device; a geometric model acquisition module is configured to obtain a high-precision three-dimensional geometric model of the green body by surface topography measurement of the green body to be sintered by a 3D optical scanning device; a space model acquisition module is configured to obtain a green body fusion defect space model by performing geometric distortion correction after spatial alignment and fusion of the three-dimensional density distribution map of the green body and the high-precision three-dimensional geometric model of the green body; a defect detection module is configured to obtain a plurality of defect space distribution characteristics corresponding to a plurality of defect types by performing fine-grained defect detection on the green body fusion defect space model; a correction parameter output module is configured to output a plurality of single-dimensional pressing correction parameters by performing pressing associated defect inversion according to the plurality of defect space distribution characteristics; a coupling parameter output module is configured to output a global pressing coupling parameter by performing spatial multivariate coupling processing of the plurality of single-dimensional pressing correction parameters based on the spatial attributes of the green body; and a batch pressing module is configured to perform batch pressing of the green body to be sintered by using the global pressing coupling parameter.

[0008] One or more technical solutions provided in the present application have at least the following technical effects or advantages: The three-dimensional density distribution diagram of the sintering silicon carbide blank is obtained by tomography of the sintering silicon carbide blank to be sintered through an industrial CT device; the surface topography of the sintering silicon carbide blank is measured through a 3D optical scanning device to obtain a high-precision three-dimensional geometric model of the sintering silicon carbide blank; after fusing the three-dimensional density distribution diagram of the sintering silicon carbide blank and the high-precision three-dimensional geometric model of the sintering silicon carbide blank, geometric distortion correction is performed to obtain a sintering silicon carbide blank fusion defect space model; fine-grained defect detection is performed to obtain a plurality of defect space distribution characteristics corresponding to a plurality of defect types; pressing associated defect inversion is performed to output a plurality of single-dimensional pressing correction parameters; spatial multivariate coupling processing of the plurality of single-dimensional pressing correction parameters is performed to output a global pressing coupling parameter; and the global pressing coupling parameter is used for batch pressing of the sintering silicon carbide blank to be sintered. The technical effect of realizing online detection of sintering silicon carbide blank defects and optimization of pressing parameters is achieved, and the pressing quality and consistency are improved. BRIEF DESCRIPTION OF DRAWINGS

[0009] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0010] Figure 1 The pressing parameter optimization method based on online detection of sintering silicon carbide blank defects provided by the embodiments of the present application is shown in the flowchart.

[0011] Figure 2 The pressing parameter optimization system structure based on online detection of sintering silicon carbide blank defects provided by the embodiments of the present application is shown in the schematic diagram.

[0012] Legend: distribution diagram acquisition module 10, geometric model acquisition module 20, space model acquisition module 30, defect detection module 40, correction parameter output module 50, coupling parameter output module 60, batch pressing module 70. DETAILED DESCRIPTION

[0013] The present application provides a pressing parameter optimization method and system based on online detection of sintering silicon carbide blank defects, which is used to solve the technical problems of low detection accuracy of sintering silicon carbide blank defects, lack of pertinence in pressing parameter optimization, and large quality fluctuation of batch pressed blanks in the prior art.

[0014] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0015] Example 1, as Figure 1 As shown, this application provides a method for optimizing pressing parameters based on online detection of defects in silicon carbide preforms. The method includes: Step S100: Perform a tomographic scan of the silicon carbide blank to be sintered using an industrial CT scanner to obtain a three-dimensional density distribution map of the blank.

[0016] Specifically, the industrial CT equipment is used to perform a comprehensive tomographic scan of the silicon carbide blank to be sintered. Utilizing the characteristics of industrial CT rays, which can penetrate the blank and generate different attenuation signals based on the material density differences in different areas of the blank, density data of the blank at various depths from the surface to the interior are collected. Subsequently, the multiple sets of continuous cross-sectional density data obtained from the scan are integrated and reconstructed using three-dimensional reconstruction technology. Finally, a three-dimensional density distribution map of the blank is generated, which can accurately reflect the overall density distribution inside the silicon carbide blank to be sintered. This distribution map will serve as the key internal information foundation for subsequent spatial alignment and fusion with the high-precision three-dimensional geometric model of the blank, and for constructing a spatial model of the blank's fusion defects. It provides core data support for the internal density dimension of the blank for subsequent fine-grained defect detection.

[0017] Step S200: The surface morphology of the silicon carbide blank to be sintered is measured by a 3D optical scanning device to obtain a high-precision three-dimensional geometric model of the blank.

[0018] Specifically, 3D optical scanning equipment, such as structured light scanning equipment or laser scanning equipment, is used to perform a full-surface coverage scan of the silicon carbide blank to be sintered. Utilizing the principle of reflection / diffuse reflection between the optical signals emitted by the equipment and the blank surface, millions or even tens of millions of feature point coordinate data are collected on the blank surface to form raw point cloud data containing the blank surface contour, texture, and dimensional details. Next, preprocessing operations are performed on the raw point cloud data. Noise reduction algorithms, such as statistical filtering and radius filtering, are used to remove abnormal points caused by environmental interference during the scanning process. Then, point cloud registration and stitching techniques are used to integrate the local point cloud data obtained from scanning the blank at different angles into a complete global point cloud model. Finally, based on the processed complete point cloud data, three-dimensional meshing and model reconstruction algorithms, such as the Poisson surface reconstruction algorithm, are used to construct a high-precision three-dimensional geometric model of the blank that is highly consistent with the actual external geometry of the blank and whose dimensional accuracy meets the requirements of subsequent spatial registration and geometric verification.

[0019] Step S300: After the space alignment fusion of the blank three-dimensional density distribution map and the blank high-precision three-dimensional geometric model, geometric distortion correction is performed to obtain a blank fusion defect space model.

[0020] Specifically, the first natural feature point distribution and the second natural feature point distribution are identified in the blank three-dimensional density distribution map and the blank high-precision three-dimensional geometric model, respectively. An iterative closest point algorithm is used to register the two sets of natural feature point distributions and construct a space transformation matrix. The two models are mapped to a unified coordinate system according to the matrix to complete space registration and obtain a preliminary registration model. Then, the preliminary registration model is segmented into multiple grid-level spatial pixel points according to a preset grid scale. The position and shape distortion of each grid-level spatial pixel point are corrected based on the optical geometric true value corresponding to the blank high-precision three-dimensional geometric model, and multiple corrected spatial pixel points are generated. Finally, all corrected spatial pixel points are re-integrated by a reverse recombination algorithm to restore and output a blank fusion defect space model that combines the internal density information and external geometric information of the blank.

[0021] Step S400: Perform fine-grained defect detection on the blank fusion defect space model to obtain multiple defect spatial distribution characteristics corresponding to multiple defect types.

[0022] Specifically, according to the multiple defect size distribution sets corresponding to the multiple defect types, the multi-scale defect characteristics are analyzed and output. Then, based on the multi-scale defect characteristics, the blank fusion defect space model is subjected to multi-scale spatial segmentation to generate a series of scale analysis models. Next, multiple defect original image sets and multiple defect shape identification sets of multiple defect types are called in a network and used as training data to perform multi-directional parameter optimization on the basic defect detection model, construct multiple single-task defect detection modules, and form a multi-defect parallel detection network in a logically isolated manner. The series of scale analysis models are input into the network one by one to perform defect parallel reasoning detection and output a series of scale intermediate results. Then, the series of scale intermediate results are decomposed according to defect types to obtain multiple series of multi-scale pixel-level defects. According to the multi-scale defect characteristics, multi-scale defect confidence values are set, multiple defect probability maps are generated for the spatially aligned multiple series of multi-scale pixel-level defects, and weighted probability fusion is performed based on multi-scale defect confidence value mapping to output multiple initial defect spatial distributions. Finally, based on the blank high-precision three-dimensional geometric model, geometric distortion verification is performed on the multiple initial defect spatial distributions, and finally multiple defect spatial distribution characteristics corresponding to multiple defect types are output, providing accurate defect information basis for subsequent pressing associated defect inversion.

[0023] Step S500: Perform pressing associated defect inversion according to the multiple defect spatial distribution characteristics to output multiple single-dimensional pressing correction parameters.

[0024] Specifically, for each defect spatial distribution feature, a corresponding defect feature vector is quantitatively extracted, which needs to cover key information such as defect type, defect volume, defect density, defect spatial orientation, and defect structure position distribution. For example, a first defect feature vector containing the above dimensions is extracted from the first defect spatial distribution feature. Then, the extracted defect feature vector is input into a pre-constructed defect pressing correlation rule library for matching. The rule library is formed by first performing process simulation repair to record a plurality of sample defect vectors corresponding to a plurality of sample pressing correction parameters, then performing feature space enhancement on the sample data, and finally constructing the causal correlation of sample defect vectors and sample pressing correction parameters through a decision tree. Through matching, single-dimensional pressing correction parameters corresponding to each defect feature vector are output, including but not limited to pressing pressure gradient adjustment amount, holding time correction amount, and demolding rate optimization value. Finally, a plurality of single-dimensional pressing correction parameters are obtained, providing a basis for subsequent spatial multivariate coupling processing.

[0025] Step S600: Spatial multivariate coupling processing of the plurality of single-dimensional pressing correction parameters based on the dough space attribute, outputting a global pressing coupling parameter.

[0026] Specifically, based on the dough space attribute, key spatial features such as the shape, size, and material distribution of different regions of the dough are covered to construct a process sensitivity field model that can reflect the sensitivity of different spatial positions of the dough to changes in pressing process parameters. Subsequently, with dough density optimization and density distribution uniformization as the core multi-objective, the output multiple single-dimensional pressing correction parameters such as pressing pressure gradient adjustment amount, holding time correction amount, and demolding rate optimization value are mapped into the constructed process sensitivity field model. Through spatial multivariate coupling optimization algorithm, the mutual influence and adaptation relationship of each single-dimensional pressing correction parameter in different dough spatial regions is coordinated, and the possible conflicts between parameters are eliminated, finally outputting a global pressing coupling parameter that can meet the process requirements of each spatial region of the dough and satisfy the multi-objective optimization requirements, providing unified and precise process parameter guidance for subsequent batch pressing of the sintered silicon carbide dough.

[0027] Step S700: Batch pressing of the sintered silicon carbide dough using the global pressing coupling parameter.

[0028] Specifically, the output global pressing coupling parameters are transmitted to a batch pressing equipment control system of the sintered silicon carbide green body to be sintered, and then the pressing equipment automatically adjusts core process indicators in the pressing process according to the global pressing coupling parameters, including but not limited to pressing pressure gradient, pressure holding time, demolding rate and the like corresponding to the parameters, and finally, the sintered silicon carbide green body is subjected to standardized batch pressing operation according to the adjusted process parameters, so as to ensure that the batch-produced green body can avoid the defect types detected in the early stage, reduce the overall defect rate, and at the same time ensure that the key quality indicators such as green body density and geometric shape meet the production requirements, thereby realizing efficient and high-quality batch production based on online defect detection and parameter optimization.

[0029] In one possible implementation, step S300 further includes: Step S310: mapping the green body three-dimensional density distribution map and the green body high-precision three-dimensional geometric model to a unified coordinate system to perform spatial registration, to obtain a preliminary registration model.

[0030] Step S320: after the preliminary registration model is segmented into a plurality of grid-level spatial pixel points according to a preset grid scale, distortion correction is performed on the plurality of grid-level spatial pixel points based on optical geometric true values, to obtain a plurality of corrected spatial pixel points.

[0031] Step S330: reversely recombining the plurality of corrected spatial pixel points to restore and output the green body fusion defect space model.

[0032] Specifically, first, first natural feature point distribution is identified in the green body three-dimensional density distribution map, and these feature points are usually selected from regions with obvious density mutations, such as internal pore edges and density anomaly boundaries, to ensure that they have clear spatial recognition; meanwhile, second natural feature point distribution is identified in the green body high-precision three-dimensional geometric model, and positions with significant geometric shapes are preferentially selected, such as surface protrusion vertices, process hole edges and corner intersection points, to ensure that the two types of feature points can form corresponding associations. Then, an iterative closest point algorithm is used to perform registration calculation on the first natural feature point distribution and the second natural feature point distribution: the closest point pairs in the two types of feature point sets are found through continuous iteration, the spatial error between the point pairs is calculated, the spatial pose of the green body three-dimensional density distribution map is adjusted, and the adjustment is continued until the overall error is reduced to below a preset threshold, and in this process, a spatial transformation matrix that can realize spatial conversion of the two types of data is constructed. Finally, according to the spatial transformation matrix, all spatial coordinates of the green body three-dimensional density distribution map are mapped to a unified coordinate system with the green body high-precision three-dimensional geometric model as a reference, so that the two are accurately aligned in spatial position, while the density data of the green body three-dimensional density distribution map and the geometric shape data of the green body high-precision three-dimensional geometric model are retained, and finally a preliminary registration model is fused to lay a spatial consistency foundation for subsequent geometric distortion correction.

[0033] According to the actual size of the silicon carbide blank to be sintered and the defect detection accuracy requirement, a preset grid scale is determined, which needs to consider the detection accuracy and data processing efficiency to ensure that both the details of the small defects can be captured and the data volume is not too large and the processing time is not too long due to too fine grid. Then, taking the preset grid scale as the segmentation standard, the obtained preliminary registration model is discretized and segmented in three-dimensional space, and the continuous preliminary registration model is uniformly divided into a large number of non-overlapping grid-level spatial pixel points, each pixel point corresponds to a unique spatial coordinate and the blank density data at that position, forming a discretized pixel point set. Next, the optical geometric true value verified by 3D optical scanning in the high-precision three-dimensional geometric model of the blank is called, which includes the accurate geometric coordinates of the surface and internal key structures of the blank, such as surface contour points, internal process hole center coordinates, etc. These coordinates have been calibrated through multiple scans and have very high accuracy. Then, for each grid-level spatial pixel point, its current spatial coordinate is extracted and compared with the standard coordinate at the corresponding position in the optical geometric true value, and the deviation value in the X, Y and Z coordinate axis directions, i.e. distortion error, is calculated. According to the calculated distortion error, a polynomial fitting correction algorithm is used to correct the coordinates of each grid-level spatial pixel point: by constructing an error fitting function, the deviation value is converted into a coordinate correction amount, the original coordinates of the pixel point are adjusted, and the position deviation caused by industrial CT scanning distortion, spatial registration deviation and other factors is eliminated. Finally, after the distortion correction of all grid-level spatial pixel points is completed, a plurality of corrected spatial pixel points with accurate position and high matching of coordinates and optical geometric true value are formed.

[0034] A core benchmark for reverse recombination is determined, a unified coordinate system is set as a spatial framework, and all corrected spatial pixel points are ensured to maintain coordinate consistency during the recombination process. Subsequently, the obtained multiple corrected spatial pixel points, each containing accurate spatial coordinates and blank density data at the corresponding position, are sorted in a three-dimensional grid according to the spatial coordinates of the pixel points, ensuring that each pixel point can be returned to the original spatial position corresponding relationship before the preliminary registration model is segmented. Then, a three-dimensional space interpolation and voxel reconstruction algorithm is used to continuously process the discrete corrected spatial pixel points: the density information at the gap between the pixel points is supplemented by calculating the density gradient between adjacent pixel points, and the model fault caused by discrete segmentation is eliminated; at the same time, based on the coordinate position of the pixel points, a continuous three-dimensional voxel structure is constructed, so that the reconstructed model can not only completely retain the external geometric shape of the blank and be consistent with the blank high-precision three-dimensional geometric model, but also accurately present the internal density distribution difference. During the recombination process, the geometric integrity and density data accuracy of the model need to be checked in real time, and the key dimensions of the reconstructed model, such as length, diameter, and corresponding dimensions of the blank high-precision three-dimensional geometric model, are compared to ensure that the error is within the allowable range; at the same time, the density values of the internal key areas are randomly checked to verify whether they are consistent with the original data of the blank three-dimensional density distribution map. Finally, after the reverse recombination of all corrected spatial pixel points is completed, the blank fusion defect space model with accurate geometric shape and real internal density distribution is output, which can be directly used for subsequent fine-grained defect detection and provide complete and accurate three-dimensional data support for identifying internal pores, surface cracks and other defects of the blank.

[0035] In one possible implementation manner, the step S310 further includes: Step S311: Identifying a first natural feature point distribution and a second natural feature point distribution from the blank three-dimensional density distribution map and the blank high-precision three-dimensional geometric model, respectively.

[0036] Step S312: Registering the first natural feature point distribution and the second natural feature point distribution by using an iterative closest point algorithm to construct a spatial transformation matrix.

[0037] Step S313: Mapping the blank three-dimensional density distribution map and the blank high-precision three-dimensional geometric model to a unified coordinate system to perform spatial registration according to the spatial transformation matrix, and obtaining the preliminary registration model.

[0038] Specifically, for the green body three-dimensional density distribution map, a density gradient threshold segmentation algorithm combined with a feature point extraction algorithm is used to identify the first natural feature point distribution: by setting a reasonable density gradient threshold, regions with sudden changes in internal density of the green body are screened out, such as the junction of internal pores and dense regions, the edge of local density abnormal blocks, etc., and then corner detection and contour sampling are performed on these regions to extract points with spatial uniqueness as the first natural feature points, and finally a first natural feature point distribution containing multiple such points is formed, ensuring that each point can accurately reflect the key position of the internal density structure of the green body. At the same time, for the high-precision three-dimensional geometric model of the green body, an edge detection algorithm and a geometric feature recognition algorithm are used to identify the second natural feature point distribution: first, the contour lines of the green body surface are extracted through edge detection, such as the edges of the green body, the circumferential edges of the process holes, etc., and then geometrically significant points are selected on the contour lines, such as the vertexes of the edges, the centers of the process holes, the highest points of the surface protrusions, etc., which can clearly represent the key positions of the external geometric structure of the green body, and finally the second natural feature point distribution is integrated to form a potential corresponding relationship between the two types of natural feature points in the spatial position, laying a foundation for subsequent spatial registration.

[0039] The identified first natural feature point distribution and second natural feature point distribution are imported into the calculation module of the iterative closest point algorithm, the spatial correspondence relationship of the two sets of feature points is initialized, the Euclidean distance is taken as the measurement standard, the second natural feature point closest in distance is matched for each first natural feature point, and an initial point pair set is formed. Then, based on the initial point pair set, a preliminary spatial transformation matrix is calculated through the singular value decomposition algorithm, which contains parameters of translation, rotation and scaling in three dimensions, and can realize the spatial mapping of the first natural feature point distribution to the second natural feature point distribution. Subsequently, the preliminary spatial transformation matrix is used to perform coordinate transformation on all first natural feature points to obtain the transformed first natural feature point distribution, and the distance error of each transformed point and the corresponding second natural feature point is recalculated, and the abnormal point pairs with error exceeding the preset threshold, such as 0.01 mm, are screened out and removed, and a new nearest point pair set is updated. The iterative process of calculating the spatial transformation matrix, transforming the feature points, updating the point pairs, and calculating the error is repeatedly executed until the overall error change of two consecutive iterations is less than the set convergence threshold, at which time the algorithm converges and outputs the final spatial transformation matrix. The matrix can accurately describe the spatial conversion relationship between the green body three-dimensional density distribution map and the green body high-precision three-dimensional geometric model, and provide key parameter support for subsequent spatial registration of the two in a unified coordinate system.

[0040] A reference of the unified coordinate system is determined, usually taking the original coordinate system of the high-precision three-dimensional geometric model of the blank as the reference coordinate system. Due to the high precision of 3D optical scanning, the coordinate system can provide reliable position reference for spatial registration. Then, the constructed spatial transformation matrix is called, which contains translation, rotation and scaling parameters, and the spatial conversion of the three-dimensional density distribution map of the blank and the reference coordinate system can be realized. Next, all spatial coordinate data of the three-dimensional density distribution map of the blank are input into the coordinate transformation module, and the coordinate conversion operation is performed according to the spatial transformation matrix. The density distribution data originally in the industrial CT scanning coordinate system is accurately mapped into the unified coordinate system taking the high-precision three-dimensional geometric model of the blank as the reference, so that the internal density distribution information of the blank and the external geometric shape information form a corresponding relationship in the spatial position. In this process, the density value data of the three-dimensional density distribution map of the blank and the geometric contour data of the high-precision three-dimensional geometric model of the blank need to be preserved synchronously to ensure that the data fusion is realized in the unified coordinate system rather than data loss. Finally, the mapped three-dimensional density distribution map of the blank and the high-precision three-dimensional geometric model of the blank are spatially aligned and verified. By comparing the coordinate deviation of the key feature points, it is confirmed that the registration accuracy meets the preset requirements, such as the deviation being less than 0.02 mm. After the verification is passed, the preliminary registration model with internal density distribution and external accurate geometric shape of the blank can be output.

[0041] In one possible implementation manner, the step S400 further includes: Step S410: analyzing and outputting multi-scale defect characteristics according to the plurality of defect size distribution sets of the plurality of defect types.

[0042] Step S420: performing multi-scale spatial segmentation on the blank fusion defect spatial model by using the multi-scale defect characteristics to generate a series of scale analysis models.

[0043] Step S430: inputting the series of scale analysis models into a multi-defect parallel detection network to perform defect parallel inference detection and output a series of scale intermediate results.

[0044] Step S440: after the probability fusion of the series of scale intermediate results, performing geometric distortion verification based on the high-precision three-dimensional geometric model of the blank to output a plurality of defect spatial distribution characteristics corresponding to the plurality of defect types.

[0045] Specifically, various defect types commonly found in the production of sintered silicon carbide blanks are identified, such as internal porosity, surface cracks, local density abnormalities, etc. For each defect type, a corresponding set of defect size distributions is obtained, which includes key size data such as length, width, depth, volume, and cross-sectional dimensions of the corresponding defects under different production batches and different pressing conditions. Subsequently, statistical analysis is performed on the defect size distribution sets of each defect type. By analyzing the distribution frequency, size variation, and typical morphological characteristics of defects in different size intervals, multi-scale defect characteristics that can distinguish size differences between different defect types and adapt to subsequent multi-scale detection needs are extracted and output. For example, micron-level size features are extracted for small pores, and millimeter-level extension size features are extracted for surface cracks, providing accurate scale division basis for multi-scale spatial segmentation of the blank fusion defect space model in subsequent steps.

[0046] The output multi-scale defect characteristics, including size intervals and morphological feature parameters of different defect types, are converted into scale constraints for the neural network. Meanwhile, the blank fusion defect space model is preprocessed into three-dimensional voxel data that conforms to the network input format, which covers the internal density information and external geometric information of the blank. Then, an improved 3D U-Net multi-scale neural network is started. The network encoder part extracts multi-scale features from the voxel data using different size convolution kernels, such as 1x1x1, 3x3x3, and 5x5x5. Small size convolution kernels focus on local features of small defects such as micron-level pores, while large size convolution kernels capture global distribution features of defects such as millimeter-level cracks and centimeter-level density abnormalities. Each encoding layer outputs a feature map corresponding to the multi-scale defect characteristics. Subsequently, the network decoder part fuses the feature maps of different encoding layers through skip connections and performs hierarchical semantic segmentation on the feature maps based on the scale constraints. The segmentation is divided into three dimensions: macro, which adapts to centimeter-level density abnormalities; meso, which adapts to millimeter-level cracks; and micro, which adapts to micron-level pores. Each dimension generates a scale sub-model. Finally, the scale sub-models in each dimension are calibrated and verified for completeness to ensure that each sub-model accurately covers the defect detection needs of the corresponding scale, and an ordered series of scale analysis models is formed.

[0047] For various defect types of silicon carbide blanks, such as internal porosity, surface cracks, and local density anomalies, single-task defect detection sub-models based on CNN are constructed for each defect type. The original defect image set and defect morphology identification set are used as training data, ResNet50 is used as the basic backbone network, network parameters are initialized through transfer learning, and spatial attention modules are introduced to strengthen defect region feature extraction. Through the Adam optimizer and cross-entropy loss function, each sub-model is iteratively trained to accurately adapt to the detection task of a single defect type, forming multiple special detection sub-models. Then, a multi-defect parallel detection network is constructed by connecting these sub-models in parallel. Each sub-model receives data independently and the inference process does not interfere with each other. The network has a built-in data distribution module that can distribute each scale sub-model in the series scale analysis model to the corresponding defect scale sub-model. For example, micron-level sub-models are distributed to CNN sub-models for detecting internal porosity, and millimeter-level sub-models are distributed to CNN sub-models for detecting surface cracks. Then, start the parallel inference process. Each CNN sub-model performs feature extraction on the received scale sub-model. The convolution layer and the pooling layer extract defect texture, contour features, and feature mapping. The fully connected layer converts the features into defect probability distribution, and outputs single-scale detection results containing defect location coordinates, defect confidence, and defect morphology parameters. Finally, the network data integration module arranges all single-scale detection results output by the sub-models in order of scale to form a series of scale intermediate results covering all sub-model detection data of the series scale analysis model.

[0048] According to the determined multiple defect types, the series of scale intermediate results are classified and decomposed by defect type to obtain multiple series of multi-scale pixel-level defect data corresponding to each defect type. Then, reference multi-scale defect characteristics are set to adapt to the multi-scale defect confidence value for pixel-level defects of different scales, such as setting a large-scale defect to a high confidence value due to significant features, and setting a small-scale defect to a corresponding low confidence value. Spatial alignment operation is performed on multi-scale pixel-level defect data of the same defect type to ensure that the pixel coordinates of the same defect region at different scales match, and defect probability maps corresponding to each scale are generated. Then, a weighted coefficient matrix is constructed according to the multi-scale defect confidence value to perform weighted probability fusion on the multiple defect probability maps after spatial alignment, eliminate the deviation of different scale detection results, and output the initial defect spatial distribution of each defect type. Finally, the obtained high-precision three-dimensional geometric model of the blank is taken as the reference to extract the optical geometric true value in the model, and the position, shape and size of each initial defect spatial distribution are checked for geometric distortion. The spatial offset error caused by segmentation or detection process is corrected, and finally multiple defect spatial distribution features corresponding to multiple defect types and accurate spatial information are output.

[0049] In one possible implementation, step S430 further includes: Step S431: calling a plurality of defect original image sets and a plurality of defect morphology identification sets of the plurality of defect types through networking.

[0050] Step S432: taking the plurality of defect original image sets and the plurality of defect morphology identification sets as training data, performing multi-directional parameter adjustment optimization of a basic defect detection model, and constructing a plurality of single-task defect detection modules.

[0051] Step S433: connecting the plurality of single-task defect detection modules in a logic isolation manner to construct the multi-defect parallel detection network.

[0052] Step S434: inputting a plurality of single-scale pixel-level models in a first scale analysis model into the multi-defect parallel detection network one by one, performing defect parallel inference detection, outputting a plurality of single-scale pixel-level defects, and constructing a first scale intermediate result.

[0053] Step S435: inputting the series of scale analysis models into the multi-defect parallel detection network by analogy, performing defect parallel inference detection, and outputting the series of scale intermediate results.

[0054] Specifically, a plurality of defect original image sets and a plurality of defect morphology identification sets corresponding to a plurality of defect types of a silicon carbide blank, such as internal porosity, surface cracks, local density abnormalities, etc., are called through networking, wherein the defect original image set contains high-definition three-dimensional image data of various defects under different production conditions, and the defect morphology identification set labels the category, boundary, size and morphology characteristic parameters of the corresponding defect, providing complete sample data support for subsequent model training.

[0055] The multiple defect original image sets corresponding to multiple defect types obtained by the networked call contain high-definition three-dimensional image data of various defects of the silicon carbide billet under different working conditions and multiple defect morphology identification sets containing information such as defect category, boundary, size, and morphological feature parameters. The multiple defect original image sets are preprocessed, including image denoising, pixel-level annotation calibration, data enhancement such as rotation, scaling, and mirror processing, and are divided into a training set and a verification set in a preset 8:2 ratio as training data for a basic CNN defect detection model. Then, a basic CNN model is configured for each defect type, ResNet50 is selected as the backbone network, a spatial attention module is introduced to strengthen defect region feature extraction, the training set data of a single defect type is used to initialize and train the corresponding CNN model, and multi-directional parameter adjustment optimization is carried out. The convolution kernel size is adjusted, such as setting a 3x3x3 convolution kernel for micron-level pores and a 5x5x5 convolution kernel for millimeter-level cracks to adapt to the defect scale, the learning rate is optimized, the cosine annealing strategy is adopted to gradually reduce the gradient from 0.001 to 0.0001 to improve the convergence stability of the model, and the iteration number is adjusted. The loss function value of the verification set no longer decreases as the termination condition, and 100-200 rounds are usually set to ensure sufficient training. Finally, the performance of each CNN model after parameter adjustment is verified by the verification set data, and the model with a defect recognition accuracy of ≥95% and a false detection rate of ≤3% is selected. Each performance meeting the CNN model is used as a special detection tool for the corresponding defect type, and multiple single-task defect detection modules corresponding to multiple defect types are constructed.

[0056] The multiple single-task defect detection modules are constructed, each module corresponds to a defect type of the silicon carbide billet, such as an internal pore detection module, a surface crack detection module, and a local density anomaly detection module, and each module has an independent data processing boundary. An exclusive input interface, a data cache unit, and an output channel are configured for each module to ensure that the modules do not occupy resources during data reception, inference calculation, and result output. Then, a parallel framework is built using a logical isolation architecture, and the scheduling controller is used to realize the cooperative scheduling of the single-task defect detection modules. The scheduling controller can automatically distribute model data to the single-task module corresponding to the detection scale according to the input scale analysis model type, such as a model suitable for micron-level defects and a model suitable for millimeter-level defects. At the same time, the logical isolation protocol is used to block invalid data interaction between the modules to avoid detection deviation caused by data confusion. Finally, a result temporary storage and integration unit is integrated into the framework, and the detection results output by each single-task defect detection module can be stored in the unit in real time and do not overlap with each other, ultimately forming a multiple defect parallel detection network with multiple defect parallel detection capability, independent operation of each module, and cooperative work of each module.

[0057] The properties of the first scale analysis model, which is a sub-model corresponding to a specific defect scale in the generated series scale analysis model, are determined. The first scale analysis model includes a plurality of single-scale pixel-level models divided according to a preset grid, and each single-scale pixel-level model covers the local density information and geometric characteristics of the blank at the scale. Then, the scheduling mechanism of the multi-defect parallel detection network is started, and the plurality of single-scale pixel-level models in the first scale analysis model are sequentially input into the network. The network scheduling controller automatically activates the single-task defect detection module corresponding to the defect detection task according to the scale properties of the single-scale pixel-level model. For example, the internal pore detection module is activated for the micron-level model, while other modules are kept in a ready state to avoid resource occupation. Each activated single-task defect detection module performs parallel inference detection on the input single-scale pixel-level model. The pixel-level defect features are extracted through the convolution layer and the pooling layer of the CNN network, and the attention mechanism is used to focus on the suspected defect area. A set of single-scale pixel-level defect data including the defect position coordinates, the defect confidence, and the defect morphology parameters is output. After all single-scale pixel-level models in the first scale analysis model are detected, the output multiple sets of single-scale pixel-level defect data are sorted, de-duplicated, and verified according to the input order to ensure data integrity and consistency, and finally the first scale intermediate result covering the full-range detection result of the first scale analysis model is formed.

[0058] Batch detection is performed on all remaining scale analysis models in the series scale analysis model. The scale hierarchy and corresponding defect detection requirements of the series scale analysis model are sorted out, and the number of single-scale pixel-level models, scale properties, and adaptive single-task defect detection modules included in each non-first scale analysis model are determined. Then, the plurality of single-scale pixel-level models in each scale analysis model are sequentially input into the multi-defect parallel detection network in order from fine to coarse or from coarse to fine. The network scheduling controller automatically activates the matching single-task defect detection module according to the scale properties of the current input model. Each activated module synchronously performs defect parallel inference detection on the single-scale pixel-level model and outputs single-scale pixel-level defect data including the defect position, confidence, and morphology parameters. After all single-scale pixel-level models in each scale analysis model are detected, the intermediate results corresponding to the respective scales are formed. Finally, the first scale intermediate result and the subsequent scale intermediate results are collected in order according to the scale to ensure that the detection data of all scales completely covers the detection range of the series scale analysis model, and finally the series scale intermediate result including the full-scale defect detection information is output.

[0059] In one possible implementation, step S440 further includes: Step S441: decomposing the series scale intermediate result according to the plurality of defect types to obtain a plurality of series multi-scale pixel-level defects.

[0060] Step S442: Set a multi-scale defect confidence value according to the multi-scale defect characteristics.

[0061] Step S443: After spatially aligning the multiple series of multi-scale pixel-level defects, a weighted probability fusion of the multiple defect probability maps is performed according to the multi-scale defect confidence value mapping, and multiple initial defect spatial distributions are output.

[0062] Step S444: Geometric distortion verification of the multiple initial defect spatial distributions is performed based on the green body high-precision three-dimensional geometric model, and multiple defect spatial distribution characteristics are output.

[0063] Specifically, a plurality of defect types to be detected in the production of sintered silicon carbide green bodies, such as internal porosity, surface cracks, local density abnormalities, etc., are determined, and each defect type is taken as an independent classification dimension. Then, for the output series of scale intermediate results, defect detection data covering different scale analysis models are resolved according to the core screening basis of "defect type", and all scale detection data related to a single defect type are extracted one by one from the series of scale intermediate results, for example, for the defect type of "internal porosity", the corresponding pixel-level detection data in different scale analysis models such as micrometer level and sub-millimeter level are extracted, including defect pixel position, suspected probability, and local morphology parameters, and for surface cracks, pixel-level detection data in millimeter-level and centimeter-level scale analysis models are extracted. Finally, a data set corresponding to each defect type is formed, i.e. each data set only contains the detection performance of a single defect type at the multi-scale pixel level, thereby obtaining a plurality of series of multi-scale pixel-level defects consistent with the number of defect types, ensuring that subsequent centralized processing can be carried out for multi-scale data of a single defect type.

[0064] The multi-scale defect characteristic of the output is analyzed, which includes the internal pores, surface cracks, local density abnormalities and other key features of various defect types of the silicon carbide blank at different scales. The key features include the micron-level size range of the internal pores, the millimeter-level extension range of the surface cracks, the large-scale defect with more obvious features and easy to be accurately identified, the small-scale defect with weak features and difficult to be identified, and the historical detection accuracy data. Then, based on the multi-scale defect characteristics, a multi-scale defect confidence value is set for each defect type at different detection scales. For the millimeter-level surface cracks with obvious features and high historical detection accuracy, the linear extension feature is clear, the detection accuracy is more than 95%, and a high confidence value of 0.9-0.95 is set. For the micron-level small pores with weak features and relatively low historical detection accuracy, the volume is small and easy to be confused with the background interference, the detection accuracy is about 80%-85%, and a relatively low but meets the detection accuracy requirement confidence value of 0.7-0.8 is set. Finally, a multi-scale defect confidence value system corresponding to the defect type-detection scale is formed.

[0065] The global coordinate system of the blank fusion defect space model is taken as a unified reference for the spatial alignment operation of the obtained multiple series of multi-scale pixel-level defects. Through the coordinate calibration algorithm, the pixel-level defect data of the same defect type at different scales is mapped to the same spatial coordinate system, ensuring the accurate matching of the pixel positions of the corresponding defect regions at different scales and avoiding spatial misalignment caused by scale segmentation differences. After completing the spatial alignment, the multi-scale pixel-level defect data of each defect type is converted into a visual defect probability map, in which the value of each pixel represents the probability of the existence of the corresponding defect at that position, so as to intuitively present the spatial distribution possibility of the defects at different scales. Then, the set multi-scale defect confidence value is called to construct a weighted coefficient mapping relationship corresponding to the defect probability map scale, i.e. a high confidence value corresponds to a higher weighted coefficient, and a low confidence value corresponds to an adaptive low weighted coefficient. Then, according to the mapping relationship, the weighted probability fusion calculation is performed on the multiple defect probability maps of the same defect type, the deviation and redundancy of the detection results at different scales are eliminated through weighted summation, the weight of the detection results at high confidence scales is strengthened, and the effective information in the detection results at low confidence scales is reasonably retained. Finally, the threshold screening and morphological regularization are performed on the fused probability data to eliminate the suspected interference points with a probability lower than the set threshold, and the boundary and morphology of the defects are clarified, so as to output an initial defect spatial distribution for each defect type that can integrate multi-scale detection information, clearly present the preliminary spatial position, range and morphological characteristics of the defects in the blank.

[0066] The green body high-precision three-dimensional geometric model obtained by the 3D optical scanning device is called, which contains the accurate surface profile coordinates, internal structure size and geometric topological relationship of the silicon carbide green body to be sintered, and can be used as a reference basis for geometric distortion checking. Then, the output multiple initial defect spatial distributions are compared with the green body high-precision three-dimensional geometric model in space, and the initial defect spatial distribution is completely aligned with the global coordinate system of the geometric model through coordinate matching algorithm. Then, the geometric distortion checking is carried out: on the one hand, the position rationality of the initial defect spatial distribution is checked, and whether the defect falls within the actual geometric structure range of the green body is judged, and the abnormal defect data exceeding the green body boundary is removed; on the other hand, the shape and size consistency of the defect is checked, and the length, width, depth and other parameters of the initial defect are compared with the physical size of the corresponding area in the green body high-precision three-dimensional geometric model, and the defect size distortion caused by multi-scale segmentation or detection deviation is corrected, such as the crack size exceeding the local structure thickness of the green body is corrected to the value conforming to the actual structure. At the same time, for the position offset problem that may exist in the initial defect spatial distribution, the feature mark points in the geometric model, such as the specific protrusions on the surface of the green body and the internal process holes, are taken as the reference, and the defect coordinates are fine-tuned and calibrated. After all the initial defect spatial distributions complete the above checking, the defect position, shape, size and other information after checking are integrated to form multiple defect spatial distribution characteristics that can accurately reflect the real distribution state of the defects in the green body, and each characteristic corresponds to a defect type.

[0067] In one possible implementation manner, the step S500 further includes: Step S510: quantitatively extracting a first defect feature vector from the first defect spatial distribution characteristic, wherein the first defect feature vector includes a first defect type, a first defect volume, a first defect density, a first defect spatial orientation, and a first defect structure position distribution.

[0068] Step S520: using the first defect feature vector in the defect pressing correlation rule library to match and output a first single-dimensional pressing correction parameter, wherein the first single-dimensional pressing correction parameter includes but is not limited to a pressing pressure gradient adjustment amount, a pressure holding time correction amount, and a demolding speed optimization value.

[0069] Specifically, for the first defect type, a defect classification model based on deep learning, such as an improved ResNet model, is used to identify visual features such as defect morphology and gray scale distribution in the first defect spatial distribution feature, and match them with a preset defect type library containing labels such as internal pores and surface cracks, to automatically label and output the first defect type; for the first defect volume, a three-dimensional space integration algorithm is used to convert the discrete pixel point coordinates of the defect region in the first defect spatial distribution feature into continuous space grids, and by calculating the product of the number of grid cells occupied by the defect and the volume of a single grid cell, the accurate defect volume value is obtained; for the first defect density extraction, the green body three-dimensional space is divided into several equal volume sub-regions by a region division algorithm, the number of first defects in each sub-region is counted, and then the number of defects per unit volume is calculated based on the sub-region volume, which is the first defect density; for the first defect spatial orientation, a principal component analysis algorithm is used to reduce the dimension of the three-dimensional coordinate points of the defect region, extract the long axis and short axis direction vectors of the defect, and by calculating the angle between the vectors and the high-precision three-dimensional geometric model coordinate system (X, Y, Z axes) of the green body, the first defect spatial orientation parameters are determined and output; for the first defect structure position distribution, the three-dimensional coordinates of the defect are aligned with the structure coordinates of the high-precision three-dimensional geometric model of the green body through coordinate mapping technology, and combined with the green body structure partition data such as the preset partition of edge area, center area, and process hole peripheral area, the structure area where the defect is located is determined and the position distribution information is output. Finally, the quantification results of the above five dimensions are integrated according to the preset data format to form a complete first defect feature vector.

[0070] A pre-constructed defect suppression association rule library is called, which is formed by using a decision tree algorithm to establish the causal association between the sample defect vector and the corresponding sample suppression correction parameter obtained by process simulation repair, and is stored in the library after feature space enhancement. Subsequently, the first defect feature vector obtained, including the first defect type, volume, density, spatial orientation and structure position distribution, is input into the matching system of the rule library, and a feature comparison algorithm is used to select the most suitable association rule from the rule library for each dimension parameter of the first defect feature vector. If there is a complete matching rule, the suppression correction parameter corresponding to the rule is directly extracted; if there is no complete matching item, the parameter fusion is performed on multiple rules with the highest similarity by weighted similarity calculation. Finally, the first single-dimensional suppression correction parameter matched with the first defect feature vector is output, which at least includes the suppression pressure gradient adjustment amount, such as adjusting the suppression pressure gradient of the local area from 1.2 MPa / mm to 1.5 MPa / mm for internal pore defects, the holding time correction amount, such as extending the holding time by 15 seconds to enhance the compactness of the green body and the demolding rate optimization value, such as reducing the demolding rate from 8 mm / s to 5 mm / s to reduce defects, which can be directly used for single-dimensional adjustment of the subsequent pressing process to improve or eliminate the corresponding defects.

[0071] In one possible implementation manner, the step S510 further includes: Step S511: Record a plurality of sample suppression correction parameters of a plurality of sample defect vectors by performing process simulation repair.

[0072] Step S512: Perform feature space enhancement on the plurality of sample defect vectors and the plurality of sample suppression correction parameters to obtain a sample defect vector set and a sample suppression correction set.

[0073] Step S513: Build the causal association of the sample defect vector set and the sample suppression correction set by a decision tree to construct the defect suppression association rule library.

[0074] Specifically, a simulation platform is built to adapt to the pressing process of silicon carbide blanks. This platform needs to simulate the core process parameters such as pressing pressure, holding time, demolding rate in actual production, and can digitally simulate the generation and repair process of blank defects. Then, in the simulation platform, multiple typical defect types of silicon carbide blank samples are preset, such as internal porosity, surface cracks, and local density abnormalities. For each defect type, multiple sample defect vectors containing different defect parameters such as defect volume, defect density, defect spatial orientation, and defect structure position distribution are generated to ensure that the samples cover different defect shapes and severity. Then, for each sample defect vector, the "defect-containing blank" is simulated and repaired: first, set the initial pressing parameters for simulation pressing, and observe the retention of defects; then gradually adjust the pressing pressure gradient, holding time, demolding rate and other parameters, and re-simulate and evaluate the improvement effect of the defect after each adjustment until the defect size or number is reduced to the preset qualified standard. Finally, record the pressing parameter combination corresponding to the successful repair of each sample defect, forming multiple sample pressing correction parameters corresponding to multiple sample defect vectors, providing basic data support for subsequent construction of defect pressing correlation rule base.

[0075] For the recorded multiple sample defect vectors, a combination of data augmentation and feature enhancement is used for processing: by adding a small random disturbance such as a ±5% parameter fluctuation in each defect feature dimension such as defect volume, density, spatial orientation, a large number of derived vectors similar to but not identical to the original sample defect vector features are generated, while supplementing defect vector samples under different working conditions such as different initial pressing temperatures and raw material particle sizes, enriching the coverage of defect features; the feature parameters of each sample defect vector are normalized to convert different dimension parameters to the [0, 1] interval, eliminating the influence of dimension difference on subsequent modeling. Then, for multiple sample pressing correction parameters, combined with the reasonable fluctuation range of process parameters in actual production, linear interpolation and boundary expansion method is used to supplement data, inserting multiple intermediate parameter values uniformly within the existing parameter interval, while reasonably extending the boundary values of the parameters, such as extending the holding time from the existing 20-60 seconds to 15-70 seconds, to ensure the continuity and integrity of the parameter distribution; then the correction parameters are standardized to convert the parameter values to standardized data with a mean of 0 and a standard deviation of 1, improving the stability of subsequent model training. Finally, all the processed sample defect vectors are arranged in a unified format to form a sample defect vector set with regular structure and rich features; at the same time, the corresponding sample pressing correction parameters are arranged and filed to form a sample pressing correction set corresponding to the sample defect vector set.

[0076] The C4.5 decision tree algorithm suitable for feature correlation modeling is selected, the obtained sample defect vector set is taken as the model input feature, the input feature includes defect type, defect volume, defect density, defect spatial orientation, defect structure position distribution and the like dimensions, and the sample pressing correction set is taken as the model output label, the output label includes pressing pressure gradient adjustment amount, pressure holding time correction amount, demolding speed optimization value and the like parameters, so as to clearly define the modeling goal of defect feature-pressing parameter. Then, the input feature is discretized, and the continuous features such as defect volume and defect density are divided into multiple discrete categories according to reasonable intervals, so as to facilitate the branch node splitting of the decision tree; then the decision tree training is started, the algorithm automatically calculates the information gain ratio of each input feature, preferentially selects the feature with the maximum information gain ratio as the root node or intermediate node, divides the sample data into different subsets according to the feature value, and gradually constructs the branch structure, for example, first taking the defect type as the primary splitting feature to divide the sample into internal pores, surface cracks and the like subsets, then taking the defect volume as the secondary feature to further split each subset, until the output label, that is, the pressing correction parameter, tends to be consistent or reaches the preset tree depth threshold. During the training process, the pruning parameters of the decision tree are adjusted through cross-validation to avoid overfitting of the model and ensure that the correlation rules have good generalization ability. After the decision tree model is trained, the path from the root node to the leaf node in the tree is extracted, the feature judgment conditions on the path constitute the 'cause', and the pressing correction parameters corresponding to the leaf node constitute the 'effect', and these cause-effect correlations are arranged as structured rules, for example, if the defect type is internal pores and the volume is less than 5 cubic millimeters, the pressing pressure gradient adjustment amount is positive 0.3 megapascal per millimeter, and the pressure holding time correction amount is positive 10 seconds. Finally, the effectiveness of all rules is verified, the rules with conflicts are removed, the rules in the edge scenarios are supplemented, and then the rules are classified and archived according to the defect type and parameter category, to form a defect pressing correlation rule library which can be directly used for subsequent defect inversion matching.

[0077] In one possible implementation manner, the step S600 further includes: Step S610: constructing a process sensitivity field model based on the green body spatial attributes.

[0078] Step S620: mapping the plurality of single-dimensional pressing correction parameters to the process sensitivity field model for spatial coupling optimization with density optimization and distribution uniformization as multiple objectives, and outputting the global pressing coupling parameters.

[0079] Specifically, the core components of the green billet's spatial attributes are clearly defined, including its three-dimensional dimensions (length, width, and height), the density of raw material particles in different regions, and the structural differences between the surface and interior, such as edge rounding and internal process channels. These parameters are obtained by integrating data from previous industrial CT scans and 3D optical scans. Subsequently, a process sensitivity field model is constructed using a combination of finite element analysis and process simulation: the green billet space is divided into multiple discrete spatial units. For each unit, the influence of different pressing parameters, such as pressure and holding time, on the density distribution of that unit is simulated, quantifying the sensitivity coefficients of each unit to the pressing parameters. These sensitivity coefficients are then arranged in a matrix according to the spatial unit positions, forming a process sensitivity field model indexed by spatial coordinates and containing sensitivity information.

[0080] First, a multi-objective optimization function is set, with the overall density of the green billet reaching a preset optimal threshold, such as above 95% of the theoretical density, as the density optimization objective. The standard deviation of the density values ​​of each spatial unit is less than a preset range, such as ±2%, as the distribution uniformity objective. A dual-objective optimization function is constructed, and weighting coefficients are determined, adjusted according to the actual production requirements for density and uniformity. Next, multiple single-dimensional pressing correction parameters, such as pressing pressure gradient adjustment and holding time correction, are used as input variables and mapped to the constructed process sensitivity field model. A multivariate coupling algorithm, such as a non-dominated sorting genetic algorithm, is used to iteratively optimize the input variables: in each iteration, the simulated density values ​​of each unit of the green billet under different parameter combinations are calculated based on the process sensitivity field model, substituted into the multi-objective optimization function to calculate the fitness, and parameter combinations that balance optimal density and uniform distribution are selected. After multiple iterations until the optimization results converge, i.e., the fitness change between two consecutive iterations is less than a preset threshold, the final output is a global pressing coupling parameter that meets the dual-objective requirements. This parameter enables overall optimization of the green billet pressing process, ensuring the consistency of green billet quality in batch pressing.

[0081] Example 2 is based on the same inventive concept as the online detection method for pressing parameters based on silicon carbide billet defects in the previous examples, such as... Figure 2 As shown, this application provides a pressing parameter optimization system based on online detection of defects in silicon carbide preforms. The system and method embodiments in this application are based on the same inventive concept. The system includes: The distribution map acquisition module 10 is used to perform tomographic scanning of the silicon carbide blank to be sintered using industrial CT equipment to obtain a three-dimensional density distribution map of the blank.

[0082] The geometric model acquisition module 20 is used to measure the surface morphology of the silicon carbide blank to be sintered using a 3D optical scanning device to obtain a high-precision three-dimensional geometric model of the blank.

[0083] The space model obtaining module 30 is configured to obtain a green body fusion defect space model by performing geometric distortion correction after aligning and fusing the green body three-dimensional density distribution map and the green body high-precision three-dimensional geometric model.

[0084] The defect detection module 40 is configured to perform fine-grained defect detection on the green body fusion defect space model to obtain a plurality of defect space distribution characteristics corresponding to a plurality of defect types.

[0085] The correction parameter output module 50 is configured to perform pressing associated defect inversion according to the plurality of defect space distribution characteristics to output a plurality of single-dimensional pressing correction parameters.

[0086] The coupling parameter output module 60 is configured to perform spatial multi-variable coupling processing on the plurality of single-dimensional pressing correction parameters based on green body space attributes to output global pressing coupling parameters.

[0087] The batch pressing module 70 is configured to perform batch pressing of the to-be-sintered silicon carbide green body by using the global pressing coupling parameters.

[0088] Further, the system is further configured to implement the following functions: According to a plurality of defect size distribution sets of the plurality of defect types, a plurality of defect size distribution sets of the plurality of defect types are analyzed to output multi-scale defect characteristics; the multi-scale defect characteristics are used to perform multi-scale spatial segmentation on the green body fusion defect space model to generate a series of scale analysis models; the series of scale analysis models are input into a multi-defect parallel detection network to perform defect parallel reasoning detection to output a series of scale intermediate results; after probability fusion of the series of scale intermediate results, geometric distortion verification is performed based on the green body high-precision three-dimensional geometric model to output a plurality of defect space distribution characteristics corresponding to the plurality of defect types.

[0089] Further, the system is further configured to implement the following functions: The plurality of defect original image sets and the plurality of defect morphology identification sets of the plurality of defect types are called in a network; the plurality of defect original image sets and the plurality of defect morphology identification sets are used as training data to perform multi-directional parameter optimization of a basic defect detection model, and a plurality of single-task defect detection modules are constructed; the plurality of single-task defect detection modules are connected in parallel in a logical isolation manner to construct the multi-defect parallel detection network; a plurality of single-scale pixel-level models in a first scale analysis model are input into the multi-defect parallel detection network one by one to perform defect parallel reasoning detection to output a plurality of single-scale pixel-level defects to form a first scale intermediate result; by analogy, the series of scale analysis models are input into the multi-defect parallel detection network to perform defect parallel reasoning detection to output the series of scale intermediate results.

[0090] Further, the system is further configured to implement the following functions: The series of intermediate results are decomposed according to the plurality of defect types to obtain a plurality of series of multi-scale pixel-level defects; multi-scale defect confidence values are set according to the multi-scale defect characteristics; the plurality of series of multi-scale pixel-level defects are spatially aligned to obtain a plurality of defect probability maps, and then weighted probability fusion of the plurality of defect probability maps is performed according to the multi-scale defect confidence value mapping, and a plurality of initial defect spatial distributions are output; geometric distortion verification of the plurality of initial defect spatial distributions is performed based on the green body high-precision three-dimensional geometric model, and the plurality of defect spatial distribution characteristics are output.

[0091] Further, the system is also used to implement the following functions: A first defect feature vector is quantitatively extracted from the first defect spatial distribution characteristic, wherein the first defect feature vector includes a first defect type, a first defect volume, a first defect density, a first defect spatial orientation, and a first defect structure position distribution; the first defect feature vector is used in a defect pressing correlation rule library to match and output a first single-dimensional pressing correction parameter, wherein the first single-dimensional pressing correction parameter includes but is not limited to a pressing pressure gradient adjustment amount, a pressure holding time correction amount, and a demolding speed optimization value.

[0092] Further, the system is also used to implement the following functions: A plurality of sample pressing correction parameters of a plurality of sample defect vectors are recorded by performing process simulation repair; feature space enhancement is performed on the plurality of sample defect vectors and the plurality of sample pressing correction parameters to obtain a sample defect vector set and a sample pressing correction set; a causal correlation between the sample defect vector set and the sample pressing correction set is constructed by a decision tree to construct the defect pressing correlation rule library.

[0093] Further, the system is also used to implement the following functions: The green body three-dimensional density distribution map and the green body high-precision three-dimensional geometric model are mapped to a unified coordinate system to perform spatial registration to obtain a preliminary registration model; after the preliminary registration model is segmented into a plurality of grid-level spatial pixel points according to a preset grid scale, distortion correction of the plurality of grid-level spatial pixel points is performed based on optical geometric true values to obtain a plurality of corrected spatial pixel points; the plurality of corrected spatial pixel points are inversely recombined to restore and output the green body fusion defect spatial model.

[0094] Further, the system is also used to implement the following functions: The first natural feature point distribution and the second natural feature point distribution are respectively identified according to the green body three-dimensional density distribution map and the green body high-precision three-dimensional geometric model; the first natural feature point distribution and the second natural feature point distribution are registered by using an iterative closest point algorithm to construct a spatial transformation matrix; and the green body three-dimensional density distribution map and the green body high-precision three-dimensional geometric model are mapped to a unified coordinate system to perform spatial registration according to the spatial transformation matrix, so as to obtain the preliminary registration model.

[0095] Further, the system is also used to realize the following functions: A process sensitivity field model is constructed based on the green body spatial attributes; density optimization and distribution uniformization are taken as multi-objectives, the plurality of single-dimensional pressing correction parameters are mapped to the process sensitivity field model to perform spatial coupling optimization, and the global pressing coupling parameters are output.

[0096] It should be noted that the above-mentioned sequence of the embodiments of the present application is only for description, and does not represent the advantages and disadvantages of the embodiments. And the above-mentioned describes the specific embodiments of the present application. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multi-task processing and parallel processing are also possible or can be advantageous.

[0097] The above only describes the preferred embodiments of the present application and does not limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

[0098] The present application and the drawings are only exemplary descriptions of the present application, and are considered to cover any and all modifications, changes, combinations or equivalents within the scope of the present application. Obviously, those skilled in the art can make various modifications and changes to the present application without departing from the scope of the present application. Thus, if these modifications and changes of the present application belong to the scope of the present application and its equivalents, the present application intends to include these modifications and changes.

Claims

1. A method for optimizing pressing parameters based on online detection of defects in silicon carbide preforms, characterized in that, The method includes: A three-dimensional density distribution map of the silicon carbide billet to be sintered was obtained by performing a tomographic scan using an industrial CT scanner. The surface morphology of the silicon carbide blank to be sintered is measured by a 3D optical scanning device to obtain a high-precision three-dimensional geometric model of the blank. After spatially aligning and fusing the three-dimensional density distribution map and the high-precision three-dimensional geometric model of the blank, geometric distortion correction is performed to obtain a spatial model of the blank fusion defect. Fine-grained defect detection is performed on the blank fusion defect space model to obtain multiple defect space distribution features corresponding to various defect types; Based on the spatial distribution characteristics of the multiple defects, the associated defects are suppressed and inverted, and multiple single-dimensional suppression correction parameters are output. Based on the spatial properties of the green blank, the spatial multivariate coupling processing of the multiple single-dimensional pressing correction parameters is performed to output the global pressing coupling parameters. The global pressing coupling parameters are used to perform batch pressing of the silicon carbide blank to be sintered.

2. The pressing parameter optimization method based on online defect detection of silicon carbide preform as described in claim 1, characterized in that, Fine-grained defect detection is performed on the green blank fusion defect spatial model to obtain multiple defect spatial distribution features corresponding to various defect types. The method includes: Based on the multiple defect size distribution sets of the various defect types, the multi-scale defect characteristics are analyzed and output. Using the aforementioned multi-scale defect characteristics, the green blank fusion defect spatial model is segmented into multiple scales to generate a series of scale analysis models. The serialized scaling analysis model is input into a multi-defect parallel detection network to perform parallel defect inference detection and output serialized scaling intermediate results. After probabilistically fusing the serialized intermediate results of spatial alignment, geometric distortion verification is performed based on the high-precision three-dimensional geometric model of the blank, and multiple spatial distribution features of defects corresponding to the various defect types are output.

3. The pressing parameter optimization method based on online defect detection of silicon carbide preform as described in claim 2, characterized in that, The method involves inputting the serialized scaling analysis model into a multi-defect parallel detection network, performing parallel defect inference detection, and outputting serialized scaling intermediate results. The network accesses multiple original image sets and multiple defect morphology identifier sets for the various defect types. Using the original image sets of multiple defects and the multiple defect morphology identifier sets as training data, multi-directional parameter tuning and optimization of the basic defect detection model are performed to construct multiple single-task defect detection modules. The multiple single-task defect detection modules are connected in parallel using a logical isolation method to construct the multi-defect parallel detection network. The multiple single-scale pixel-level models in the first-scale analysis model are input one by one into the multi-defect parallel detection network to perform parallel inference detection of defects and output multiple sets of single-scale pixel-level defects to form the intermediate results of the first scale. By analogy, the serialized scaling analysis model is input into the multi-defect parallel detection network to perform parallel inference detection of defects and output the serialized scaling intermediate results.

4. The pressing parameter optimization method based on online defect detection of silicon carbide preform as described in claim 2, characterized in that, After probabilistically fusing the serialized intermediate results of spatial alignment, geometric distortion verification is performed based on the high-precision three-dimensional geometric model of the blank, and multiple spatial distribution features of defects corresponding to the various defect types are output. The method includes: Based on the various defect types, the serialized scale intermediate results are decomposed to obtain multiple serialized multi-scale pixel-level defects; Based on the aforementioned multi-scale defect characteristics, a multi-scale defect confidence value is set; After spatially aligning the multiple serialized multi-scale pixel-level defects to obtain multiple defect probability maps, the multiple defect probability maps are weighted and probabilistically fused based on the multi-scale defect confidence value mapping to output multiple initial defect spatial distributions. Based on the high-precision three-dimensional geometric model of the blank, the geometric distortion of the spatial distribution of the multiple initial defects is verified, and the spatial distribution characteristics of the multiple defects are output.

5. The pressing parameter optimization method based on online defect detection of silicon carbide preform as described in claim 1, characterized in that, Based on the spatial distribution characteristics of the multiple defects, suppression-related defect inversion is performed, and multiple one-dimensional suppression correction parameters are output. The method includes: The first defect feature vector is extracted from the spatial distribution features of the first defect, wherein the first defect feature vector includes the first defect type, the first defect volume, the first defect density, the first defect spatial orientation, and the first defect structural location distribution. Using the first defect feature vector in the defect suppression association rule base, the first single-dimensional suppression correction parameter is matched and output. The first single-dimensional suppression correction parameter includes, but is not limited to, the suppression pressure gradient adjustment amount, the holding time correction amount, and the demolding rate optimization value.

6. The pressing parameter optimization method based on online defect detection of silicon carbide preform as described in claim 5, characterized in that, The method further includes: By performing process simulation repair, multiple sample suppression correction parameters of multiple sample defect vectors are recorded; Feature space enhancement is performed on the multiple sample defect vectors and multiple sample suppression correction parameters to obtain a sample defect vector set and a sample suppression correction set. The causal relationship between the sample defect vector set and the sample suppression correction set is constructed by using a decision tree, and the defect suppression association rule base is constructed.

7. The pressing parameter optimization method based on online defect detection of silicon carbide preform as described in claim 1, characterized in that, By spatially aligning and fusing the three-dimensional density distribution map and the high-precision three-dimensional geometric model of the green blank, and then correcting for geometric distortion, a spatial model of the green blank fusion defect is obtained. The method includes: The three-dimensional density distribution map and the high-precision three-dimensional geometric model of the green blank are mapped to a unified coordinate system to perform spatial registration, and a preliminary registration model is obtained. After dividing the preliminary registration model into multiple grid-level spatial pixels according to the preset grid scale, the distortion correction of the multiple grid-level spatial pixels is performed based on the optical geometric truth to obtain multiple corrected spatial pixels. The multiple correction spatial pixels are reverse-engineered and reconstructed to restore the blank fusion defect spatial model.

8. The pressing parameter optimization method based on online defect detection of silicon carbide preform as described in claim 7, characterized in that, The method further includes: The distribution of the first natural feature point and the distribution of the second natural feature point are identified in the three-dimensional density distribution map and the high-precision three-dimensional geometric model of the green blank, respectively. The iterative nearest point algorithm is used to register the distributions of the first and second natural feature points, and a spatial transformation matrix is ​​constructed. Based on the spatial transformation matrix, the three-dimensional density distribution map of the blank and the high-precision three-dimensional geometric model of the blank are mapped to a unified coordinate system to perform spatial registration, thereby obtaining the preliminary registration model.

9. The pressing parameter optimization method based on online defect detection of silicon carbide preform as described in claim 1, characterized in that, Based on the spatial properties of the green blank, the method performs spatial multivariate coupling processing on the multiple one-dimensional pressing correction parameters to output global pressing coupling parameters. A process sensitivity field model is constructed based on the aforementioned green blank spatial properties; With density optimization and distribution homogenization as multiple objectives, the multiple one-dimensional suppression correction parameters are mapped to the process sensitivity field model for spatial coupling optimization, and the global suppression coupling parameters are output.

10. A pressing parameter optimization system based on online detection of defects in silicon carbide preforms, characterized in that, The system is used to implement the pressing parameter optimization method based on online detection of defects in silicon carbide preforms as described in any one of claims 1-9, and the system comprises: The distribution map acquisition module is used to perform tomographic scanning of the silicon carbide billet to be sintered using industrial CT equipment to obtain a three-dimensional density distribution map of the billet. The geometric model acquisition module is used to measure the surface morphology of the silicon carbide blank to be sintered using a 3D optical scanning device to obtain a high-precision three-dimensional geometric model of the blank. The spatial model acquisition module is used to obtain a spatial model of the fusion defect of the green blank by spatially aligning and fusing the three-dimensional density distribution map and the high-precision three-dimensional geometric model of the green blank, and then correcting the geometric distortion. The defect detection module is used to perform fine-grained defect detection on the blank fusion defect space model to obtain multiple defect space distribution features corresponding to various defect types. The correction parameter output module is used to perform suppression correlation defect inversion based on the spatial distribution characteristics of the multiple defects, and output multiple single-dimensional suppression correction parameters. The coupling parameter output module is used to perform spatial multivariate coupling processing on the multiple single-dimensional pressing correction parameters based on the green blank spatial properties, and output the global pressing coupling parameters. The batch pressing module is used to batch press the silicon carbide blank to be sintered using the global pressing coupling parameters.

Images