Production method of silicon carbide

Abstract

The invention relates to the field of silicon carbide production, and provides a silicon carbide production method which comprises the following steps: firstly, preparing raw materials including waste silicon carbide blocks, high-purity quartzite, waste graphite powder and the like, and then crushing and screening the raw materials to obtain pretreated raw materials suitable for subsequent processing; then the pretreated mixed raw materials are placed in a high-temperature furnace to be subjected to a carbon thermal reduction reaction to generate a reaction product, and finally the reaction product is further crushed and screened to obtain pure silicon carbide blocks. Therefore, by recycling the waste silicon carbide blocks and the waste graphite powder, effective utilization of resources can be realized, and environmental pollution is reduced.

Description

Technical Field

[0001] This application relates to the field of silicon carbide production, and more specifically, to a method for producing silicon carbide. Background Art

[0002] Silicon carbide (SiC) is an important non-oxide ceramic material. Due to its excellent mechanical properties, chemical stability, high temperature resistance, good thermal conductivity, and electrical properties, it is widely used in high-tech fields such as semiconductors, optoelectronic devices, refractory materials, abrasives, and aerospace.

[0003] Silicon carbide is mainly prepared by carbothermal reduction reaction, which involves mixing high-purity quartz sand with carbonaceous materials and reacting them at high temperature to produce silicon carbide. Traditional silicon carbide production usually relies on high-purity raw materials (such as quartz sand and graphite), while waste silicon carbide blocks and waste graphite powder are often regarded as industrial waste and not fully utilized, which causes serious waste of resources and increases the cost of waste treatment and environmental pressure.

[0004] Therefore, an optimized method for producing silicon carbide is expected. Summary of the Invention

[0005] This application aims at the deficiencies in the prior art and provides a method for producing silicon carbide.

[0006] According to one aspect of this application, a method for producing silicon carbide is provided, which includes: Step 1: Prepare raw materials, where the raw materials include waste silicon carbide blocks, high-purity quartz stones, and waste graphite powder; Step 2: Crush and screen the waste silicon carbide blocks, the high-purity quartz stones, and the waste graphite powder to obtain pre-treated raw materials; Step 3: After mixing the pre-treated raw materials, put them into a high-temperature furnace for carbothermal reduction reaction to obtain reaction products; Step 4: Crush and screen the reaction products to obtain pure silicon carbide blocks.

[0007] Due to the above technical solutions adopted in this application, significant technical effects are achieved: The method for producing silicon carbide provided in this application first prepares raw materials including waste silicon carbide blocks, high-purity quartz stones, and waste graphite powder, then crushes and screens these raw materials to obtain pre-treated raw materials suitable for subsequent processing, then places the pre-treated mixed raw materials in a high-temperature furnace for carbothermal reduction reaction to generate reaction products, and finally further crushes and screens the reaction products to obtain pure silicon carbide blocks. In this way, by recycling waste silicon carbide blocks and waste graphite powder, effective utilization of resources can be achieved and environmental pollution can be reduced. Description of the Drawings

[0008] The above and other objects, features, and advantages of the present application will become more apparent by describing the embodiments of the present application in more detail with reference to the accompanying drawings. The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation to the present application. In the drawings, the same reference numerals generally represent the same components or steps.

[0009] Figure 1 It is a flowchart of a method for producing silicon carbide according to an embodiment of the present application.

[0010] Figure 2 It is a flowchart of step 2 in the method for producing silicon carbide according to an embodiment of the present application.

[0011] Figure 3 It is a schematic diagram of data flow in step 2 of the method for producing silicon carbide according to an embodiment of the present application.

[0012] Figure 4 It is a flowchart of step 25 in the method for producing silicon carbide according to an embodiment of the present application. Detailed Description of the Embodiments

[0013] Hereinafter, exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. It should be understood that the present application is not limited by the exemplary embodiments described herein.

[0014] Silicon carbide (SiC) is an important non-oxide ceramic material. With its excellent mechanical properties, chemical stability, high-temperature resistance, as well as excellent thermal conductivity and electrical properties, it has been widely used in high-tech fields such as semiconductors, optoelectronic devices, refractory materials, abrasives, and aerospace.

[0015] The preparation of silicon carbide usually adopts the carbothermal reduction method, which uses high-purity quartz sand and carbonaceous materials as raw materials to react at high temperature to produce silicon carbide. However, the traditional silicon carbide production process highly depends on high-purity raw materials (such as quartz sand and graphite), while by-products such as waste silicon carbide blocks and waste graphite powder generated in industry have long failed to be effectively utilized and are often regarded as waste. This not only causes waste of resources, but also increases the cost of waste treatment and brings certain pressure to the environment.

[0016] Based on this, the present application proposes a method for producing silicon carbide. Figure 1 It is a flowchart of a method for producing silicon carbide according to an embodiment of the present application. As Figure 1As shown, the production method of silicon carbide according to an embodiment of the present application includes: Step 1: Prepare raw materials, where the raw materials include waste silicon carbide blocks, high-purity quartz stones, and waste graphite powder; Step 2: Crush and screen the waste silicon carbide blocks, the high-purity quartz stones, and the waste graphite powder to obtain pre-treated raw materials; Step 3: After mixing the pre-treated raw materials, put them into a high-temperature furnace for carbothermal reduction reaction to obtain a reaction product; Step 4: Crush and screen the reaction product to obtain pure silicon carbide blocks.

[0017] In Step 1, prepare raw materials, where the raw materials include waste silicon carbide blocks, high-purity quartz stones, and waste graphite powder. It should be understood that high-purity quartz stones can provide high-quality silicon sources for the synthesis of silicon carbide, ensuring the high purity and good performance of the final product, silicon carbide. Here, by using waste silicon carbide blocks and waste graphite powder as part of the raw materials, the effective reuse of resources can be achieved, reducing the dependence on primary resources and at the same time reducing production costs.

[0018] In Step 2, crush and screen the waste silicon carbide blocks, the high-purity quartz stones, and the waste graphite powder to obtain pre-treated raw materials. It should be understood that by crushing raw materials such as waste silicon carbide blocks and high-purity quartz stones, the raw material particles can be made smaller, thereby increasing their specific surface area, which is conducive to improving the efficiency of the subsequent carbothermal reduction reaction. That is, smaller raw material particles are more likely to participate in chemical reactions, helping to accelerate the reaction rate and improve the completeness of the reaction. The screening operation is to screen out particles with particle sizes that meet the requirements. Raw materials with different particle sizes perform differently in the reaction. If the particle size is too large, the reaction contact area is small and the reaction is incomplete; if the particle size is too small, the reaction may be too violent to control, and it will also increase dust pollution. A suitable particle size distribution can ensure the stable and efficient progress of the reaction, improving the yield and quality of silicon carbide. Generally speaking, through pre-treatment operations such as crushing and screening of raw materials, the raw materials can be made more uniform, reducing the problem of insufficient reaction caused by differences in particle size, and thus making the produced silicon carbide have better consistency and higher purity.

[0019] In particular, during the production process, the uniformity of the raw material particle size has an important impact on the quality of the final product. However, in the traditional process, there are no precise means of controlling the particle size during the crushing and screening of raw materials, which easily leads to non-uniform raw material particle sizes. Specifically, during the crushing process, mechanical equipment such as jaw crushers and ball mills will, due to fluctuations in equipment performance, raw material hardness, and operating conditions (such as rotation speed and time), result in a wide range of particle size distributions, generating excessive fine powder or oversized particles. The screening process usually relies on vibrating screens for screening, which is easily affected by screen blockage or unstable vibration amplitudes of the equipment, resulting in low screening accuracy. The combined effect of these factors will cause the raw materials after crushing and screening to have non-uniform particle sizes, and this non-uniformity of particle sizes will directly affect the sufficiency of the carbothermal reduction reaction, thereby leading to unstable quality of the silicon carbide product, and even problems such as high impurity content and incomplete crystallization of silicon carbide.

[0020] Based on this, in Step 2: Crushing and screening the waste silicon carbide blocks, the high-purity quartz stone, and the waste graphite powder to obtain pre-treated raw materials, the technical concept of this application is to first collect the particle size distribution image of the pre-treated raw materials by a camera, then perform multi-size screenshot processing on the particle size distribution image of the pre-treated raw materials to obtain the first and second local particle size distribution images, then extract the particle size distribution characteristics of the first and second local particle size distribution images to obtain the first and second particle size raw material distribution image characteristics maps, and finally, based on the alignment analysis representation of the raw material particle size distribution of the first and second particle size raw material distribution image characteristics maps, intelligently judge whether the particle uniformity of the pre-treated raw materials meets the preset requirements. In this way, by precisely controlling the uniformity of the raw material particle size, the sufficiency of the carbothermal reduction reaction can be significantly improved, thereby improving the crystal quality of silicon carbide, reducing the impurity content, and reducing crystallization defects.

[0021] Specifically, Figure 2 It is a flowchart of Step 2 in the production method of silicon carbide according to an embodiment of this application. Figure 3 It is a schematic diagram of data flow in Step 2 in the production method of silicon carbide according to an embodiment of this application. As Figure 2 and Figure 3As shown, step 2 includes: step 21, obtaining the particle size distribution image of the preprocessed raw material collected by the camera; step 22, cropping a first local particle size distribution image with a first size from the particle size distribution image; step 23, cropping a second local particle size distribution image with a second size from the particle size distribution image, where the second size is smaller than the first size, and the first local particle size distribution image and the second local particle size distribution image have the same shape; step 24, performing particle size distribution feature extraction on the first local particle size distribution image and the second local particle size distribution image to obtain a first particle size raw material distribution image feature map and a second particle size raw material distribution image feature map; step 25, performing a strongly causal interconnected raw material particle size distribution alignment analysis on the first particle size raw material distribution image feature map and the second particle size raw material distribution image feature map to obtain a heterogeneous particle size raw material distribution image semantic-level alignment coding feature map; step 26, based on the heterogeneous particle size raw material distribution image semantic-level alignment coding feature map, obtaining an identification result, and the identification result is used to indicate whether the particle uniformity of the preprocessed raw material meets the preset requirements.

[0022] In step 21, the particle size distribution image of the preprocessed raw material collected by the camera is obtained. It should be understood that the particle size distribution image of the preprocessed raw material mainly captures information about the size, shape, and their distribution of the raw material particles. Specifically, in the particle size distribution image of the preprocessed raw material, different-sized raw material particles can be intuitively understood. Larger particles will occupy a relatively larger pixel area in the image, while smaller particles will appear as smaller image areas. In the particle size distribution image, the number of particles of different sizes is also reflected. Particles in certain size ranges may be more numerous, showing a relatively dense distribution of particles in the corresponding size area in the image; while particles in other size ranges are fewer, appearing relatively sparse in the image. In addition to size and quantity, the shape of the raw material particles is also included in the image. The shape of the raw material particles may be diverse, such as round, oval, irregular shapes, etc. The edge features and contours of particles of different shapes are different in the image, and this shape information is also helpful for comprehensively understanding the particle size characteristics of the raw material. Generally speaking, by comprehensively considering information such as the size, shape, and distribution of the raw material particles, the overall uniformity of the preprocessed raw material particles can be understood, and thus it can be determined whether it meets the preset requirements.

[0023] The following is a detailed elaboration of a specific implementation process of "obtaining the particle size distribution image of the preprocessed raw material collected by the camera": First is the equipment selection and installation. The selection of industrial cameras is directly related to the quality of image acquisition. High-resolution cameras need to be selected to clearly present the details of raw material particles, such as the shape and edge features of tiny particles, providing a basis for subsequent accurate analysis of particle size distribution. At the same time, it should have good image capture ability to quickly and accurately freeze the instantaneous state of the raw materials, avoiding image blurring or information loss caused by shooting delays. In addition, appropriate focal length and field of view are also indispensable, which need to be determined according to the stacking method of raw materials, the size of the conveying device, and the actual production scenario to ensure that the camera can completely and clearly capture the pre-treated raw materials, making the raw materials in the image at an appropriate proportion and position for comprehensive observation and analysis of their particle size distribution. During installation, the camera should be firmly set at a position where it can clearly capture the raw materials. For example, on the raw material conveyor belt, it should be installed at an appropriate height above the conveyor belt, ensuring that the lens is perpendicular downward and maintaining an appropriate distance from the raw material surface, so that the raw materials are completely in the center of the camera's field of view. At the same time, ensure that the light evenly covers the raw material area to avoid local over-brightness or over-darkness affecting the image quality.

[0024] The setting of image acquisition parameters is an important part of this implementation process. The adjustment of the exposure time depends on the characteristics of the raw materials. If the raw materials are darker in color and absorb more light, in order to make the details of the raw material particles in the image clearly distinguishable, it is necessary to appropriately increase the exposure time; conversely, for lighter-colored or highly reflective raw materials, the exposure time should be correspondingly shortened to prevent the loss of detail information due to over-bright images. The selection of the frame rate depends on the running speed of the production line. In the case of a faster production speed, increasing the frame rate can ensure that the camera can quickly capture the continuous state of the raw materials, avoiding the omission of important information due to too long shooting intervals, and ensuring that the image can truly reflect the particle size distribution of the raw materials during the conveying process. The setting of the image resolution needs to consider multiple factors. Although a higher resolution can provide richer and more accurate details, it will greatly increase the burden of data storage and processing. Therefore, it is necessary to select an appropriate resolution on the premise of ensuring that it can meet the requirements of particle size analysis to balance the data volume and image quality.

[0025] The optimization of the acquisition environment is also crucial for obtaining high-quality particle size distribution images. The lighting condition is one of the important factors affecting the image quality. To avoid the interference of shadows and reflections on the image, a uniform and stable light source is required. The diffuse reflection light source is a good choice. It can make the light scatter gently onto the surface of the raw material, effectively reducing the generation of reflections and shadows, and making the raw material particles present clear and real contours and details in the image. Or arrange multiple light sources reasonably around the camera. By adjusting the angle and intensity of the light sources, the light can be evenly projected onto the raw material, ensuring that the entire raw material area is fully illuminated and the illumination intensity is consistent. At the same time, keeping the acquisition area clean cannot be ignored. Dust, debris, etc. may block the line of sight of the camera, resulting in problems such as blurred images and noise. Therefore, the acquisition area should be cleaned regularly to ensure the cleanliness of the environment.

[0026] Data transmission and storage are the last links in the entire image implementation process. The image data collected by the camera needs to be transmitted stably and quickly to a computer or a dedicated data storage device through a data cable (such as HDMI, etc.) or the network. During the transmission process, necessary measures should be taken to ensure the integrity and stability of the data. For example, use high-quality data cables, ensure that the interface connections are tight, and avoid data transmission interruption or loss caused by loosening; adopt appropriate network transmission protocols and optimize the network environment to prevent the impact of network fluctuations on data transmission. When storing image data, it is crucial to select a suitable file format. Formats such as JPEG and PNG have good image quality retention capabilities and also have a certain compression ratio, which can effectively save storage space without losing too much image detail. In addition, to facilitate the subsequent management, query, and analysis of the images, detailed metadata information should be added to the image data, such as the acquisition time, raw material batch number, production line number, etc. These information helps to trace and compare the particle size distribution of raw materials at different times and batches, providing strong data support for optimizing the production process.

[0027] In step 22, a first local particle size distribution image with a first size is captured from the particle size distribution image. It should be understood that the complete particle size distribution image may contain a large amount of information. Directly processing the entire image may lead to a waste of computing resources and a decrease in analysis accuracy. By capturing the first local particle size distribution image with a specific size, certain key areas can be more carefully focused on, such as the particle aggregation area, the large particle or small particle concentration area, etc. That is, by capturing the first local particle size distribution image with a first size, the particle size situation in certain special areas can be analyzed more accurately, which helps to more comprehensively understand the particle size characteristics of the raw material.

[0028] In step 23, a second local particle size distribution image with a second size is captured from the particle size distribution image, where the second size is smaller than the first size, and the first local particle size distribution image and the second local particle size distribution image have the same shape. It should be understood that by capturing a second local particle size distribution image with a second size (smaller than the first size and having the same shape as the first local particle size distribution image) from the particle size distribution image, multi-scale analysis of the raw material particles can be achieved. Specifically, the first local particle size distribution image with the first size (larger size) contains macroscopic features of the raw material particle size, such as the overall particle size distribution range, the general particle aggregation situation, etc.; while the second local particle size distribution image with the second size (smaller size) can capture the particle size details at the microscopic level, such as the morphology of small particles, the subtle differences on the particle surface, and the arrangement of particles in a small area, etc. By simultaneously obtaining local particle size distribution images of different scales, a comprehensive extraction of the particle size characteristics of the raw material from the macroscopic to the microscopic can be realized, making up for the deficiencies of single-scale analysis and helping to more comprehensively understand the true situation of the raw material particle size distribution.

[0029] In step 24, the first local particle size distribution image and the second local particle size distribution image are subjected to particle size distribution feature extraction to obtain a first particle size raw material distribution image feature map and a second particle size raw material distribution image feature map. Specifically, in the embodiment of the present application, step 24 includes: inputting the first local particle size distribution image and the second local particle size distribution image into a particle size distribution feature extractor respectively to obtain the first particle size raw material distribution image feature map and the second particle size raw material distribution image feature map. Correspondingly, in order to extract the key information related to the particle size distribution from the local particle size distribution images of different sizes obtained from the screenshots, while simplifying the data and retaining the information valuable for judging the uniformity of raw material particles, it is necessary to perform particle size distribution feature extraction on the first local particle size distribution image and the second local particle size distribution image to obtain a first particle size raw material distribution image feature map and a second particle size raw material distribution image feature map that contain information on the size, shape, and distribution characteristics of raw material particles. In a specific embodiment of the present application, the first local particle size distribution image and the second local particle size distribution image are input into a particle size distribution feature extractor respectively to obtain the first particle size raw material distribution image feature map and the second particle size raw material distribution image feature map. Particularly, the particle size distribution feature extractor is a particle size distribution feature extractor based on the Atrous Pyramid Model. Those of ordinary skill in the art should understand that the Atrous Pyramid Model can obtain receptive fields of different scales without increasing too much computational complexity through atrous convolutional layers with different dilation rates, and this model can fuse features of different scales, making full use of the information at each scale to obtain a more comprehensive and richer feature representation. Due to the multi-scale receptive field and information fusion ability of the Atrous Pyramid Model, it can more comprehensively extract the particle size features in the first local particle size distribution image and the second local particle size distribution image. Whether it is the microscopic particle details or the macroscopic particle size distribution trend, they can be effectively captured and extracted, so as to obtain a more accurate first particle size raw material distribution image feature map and a second particle size raw material distribution image feature map.

[0030] The following is a detailed elaboration of a specific implementation process of "inputting the first local particle size distribution image and the second local particle size distribution image into a particle size distribution feature extractor respectively to obtain the first particle size raw material distribution image feature map and the second particle size raw material distribution image feature map": First, a particle size distribution feature extractor needs to be prepared. Here, a particle size distribution feature extractor based on the Atrous Pyramid Model is selected. Its advantage lies in being able to capture image features at different scales, taking into account the feature extraction of both large and small particles. Before actual application, the Atrous Pyramid Model needs to be specifically designed and parameter-adjusted according to the specific requirements of the task and historical data. For example, determine the appropriate number of atrous convolutional layers and atrous rates, etc. At the same time, a large number of sample images with particle size annotation information should be used to train the feature extractor. These samples should cover various different particle size distribution situations as much as possible to improve the generalization ability of the model. During the training process, continuously adjust the model parameters so that the model can accurately extract particle size-related features from the images. After training, the model also needs to be evaluated and optimized using the validation set to ensure that its performance meets the expectations.

[0031] Before inputting the images, the first local particle size distribution image and the second local particle size distribution image need to be preprocessed. Due to differences in the shooting conditions, lighting intensity, etc. of different images, normalization processing is required to unify the pixel values of the images to a specific range, such as the [0, 1] interval, to eliminate the influence of image acquisition differences on the feature extraction results and enable the model to process different images more stably. In addition, it is necessary to ensure that the size of the input images meets the requirements of the feature extractor. If the image size is too large, it will increase the computational amount and processing time; if the size is too small, important particle size feature information may be lost. Therefore, according to the model input requirements, appropriate scaling or cropping operations should be performed on the images.

[0032] Next, enter the crucial feature extraction process. First, input the preprocessed first local particle size distribution image into the particle size distribution feature extractor, and the image starts to be processed in each layer of the model as input data. In the feature extractor, the image first passes through the atrous convolutional layer. Atrous convolution is a special convolution operation that inserts holes between the elements of the convolution kernel, which can expand the receptive field of the convolution without increasing the number of parameters. Atrous convolutional layers with different atrous rates can capture particle size features at different scales. The convolutional layer with a larger atrous rate can capture the distribution features of larger particles in the image, while the convolutional layer with a smaller atrous rate pays more attention to the detailed features of small particles. After multiple atrous convolution operations, the model generates a feature map based on the convolution results. This feature map is a numerical representation of the particle size distribution features in the input image. Different positions of each feature map correspond to the feature information of different regions in the image, and the number of channels of the feature map usually increases as the convolutional layer deepens to represent richer feature information. Then, following the same steps, input the second local particle size distribution image into the feature extractor, and after operations such as atrous convolution, generate the feature map of the second particle size raw material distribution image.

[0033] Finally, there are the output and application steps. After a series of processes, the final first-granularity raw material distribution image feature map and the second-granularity raw material distribution image feature map are obtained, and these feature maps will be saved in a specific data format for subsequent analysis and application.

[0034] In step 25, a strong causal interconnection-based raw material granularity distribution alignment analysis is performed on the first-granularity raw material distribution image feature map and the second-granularity raw material distribution image feature map to obtain a heterogeneous-granularity raw material distribution image semantic-level alignment encoded feature map. Specifically, Figure 4 is a flowchart of step 25 in the production method of silicon carbide according to an embodiment of the present application. As Figure 4 shown, the step 25 includes: step 251, performing feature decoupling on the first-granularity raw material distribution image feature map and the second-granularity raw material distribution image feature map along the channel dimension of the first-granularity raw material distribution image feature map and the second-granularity raw material distribution image feature map to obtain a set of first-granularity raw material distribution image local feature encoding vectors and a set of second-granularity raw material distribution image local feature encoding vectors; step 252, inputting the set of first-granularity raw material distribution image local feature encoding vectors and the set of second-granularity raw material distribution image local feature encoding vectors into a causal chain unit screening network to obtain a set of {first-granularity raw material distribution image local feature encoding vector, second-granularity raw material distribution image local feature encoding vector} feature pairs that constitute the granularity raw material distribution causal chain unit; step 253, inputting the set of {first-granularity raw material distribution image local feature encoding vector, second-granularity raw material distribution image local feature encoding vector} feature pairs that constitute the granularity raw material distribution causal chain unit into a causal chain unit semantic dynamic response network to obtain a set of heterogeneous-granularity raw material distribution image semantic-level alignment encoded feature vectors; step 254, performing feature aggregation on the set of heterogeneous-granularity raw material distribution image semantic-level alignment encoded feature vectors to obtain the heterogeneous-granularity raw material distribution image semantic-level alignment encoded feature map.

[0035] It should be understood that the first-granularity raw material distribution image feature map and the second-granularity raw material distribution image feature map may have significant differences in distribution, scale, and semantic background due to being derived from local images of different sizes (the first local granularity distribution image and the second local granularity distribution image). For example, the first local image feature map with a larger size may reflect the macroscopic distribution trend of the raw material granularity, while the second local image feature map with a smaller size focuses more on the microscopic particle detail features. These differences make it difficult to directly perform information interaction on the two feature maps and unable to effectively fuse the information they contain to judge the granularity uniformity of the raw material. Moreover, there may be mutual influence and causal correlation between particles of different sizes. For example, the presence of large particles may affect the distribution of small particles. Based on this, the present application performs a strong causal interconnection-based raw material granularity distribution alignment analysis on the first-granularity raw material distribution image feature map and the second-granularity raw material distribution image feature map to obtain a cross-granularity raw material distribution image semantic-level alignment encoded feature map. Specifically, the strong causal interconnection-based raw material granularity distribution alignment analysis operation can dynamically smooth the cross-domain distribution differences between the first-granularity raw material distribution image feature map and the second-granularity raw material distribution image feature map through semantic affinity by introducing a granularity raw material distribution semantic alignment modulation flow field. This means that the two feature maps can be better aligned at the semantic level, reducing the interference caused by different semantic backgrounds, and thus enabling better information interaction and fusion, thereby providing a more reliable feature representation for accurately judging the particle uniformity of the raw material.

[0036] Specifically, first, along the channel dimension of the first-granularity raw material distribution image feature map and the second-granularity raw material distribution image feature map, the first-granularity raw material distribution image feature map and the second-granularity raw material distribution image feature map are decoupled in terms of features to obtain a set of first-granularity raw material distribution image local feature encoding vectors and a set of second-granularity raw material distribution image local feature encoding vectors. The above process can be expressed by the formula: ; where represents the first-granularity raw material distribution image feature map, represents the second-granularity raw material distribution image feature map, represents the feature decoupling operation, respectively represent the 1st, 2nd, th, and th first-granularity raw material distribution image local feature encoding vectors in the set of first-granularity raw material distribution image local feature encoding vectors, respectively represent the 1st, 2nd, th, and th second-granularity raw material distribution image local feature encoding vectors in the set of second-granularity raw material distribution image local feature encoding vectors.

[0037] It should be understood that in the original first-granularity raw material distribution image feature map and the second-granularity raw material distribution image feature map, the feature elements of different channels are often highly coupled together. For example, the information about particle size may be intertwined with the information about particle shape in some channels and is difficult to distinguish. By performing feature decoupling along the channel dimension, these highly coupled feature elements can be separated. For instance, the particle size and shape features that were originally mixed in one channel can be respectively reflected in the local feature encoding vectors of different-granularity raw material distribution images after decoupling. In this way, each local feature encoding vector of the raw material distribution image of a certain granularity can more specifically represent a certain type of feature, making the representation of the features clearer and more definite. This is helpful for subsequent analysis and understanding of the raw material granularity features because only by accurately separating each feature element can we better judge their influence on the uniformity of raw material particles. And through the feature decoupling operation, it is also helpful to clarify the boundaries of the semantic representations of each channel in the original granularity raw material distribution image feature map. That is, the semantics represented by the obtained local feature encoding vectors of the raw material distribution images of each granularity are more single and definite, which is helpful for more accurately understanding and utilizing these feature information in subsequent analysis and avoiding misjudgments caused by semantic confusion.

[0038] Specifically, in the embodiment of the present application, the step 252 includes: extracting a first alternative local feature encoding vector of the raw material distribution image of a certain granularity from the set of the local feature encoding vectors of the raw material distribution image of the first granularity; extracting a second alternative local feature encoding vector of the raw material distribution image of a certain granularity from the set of the local feature encoding vectors of the raw material distribution image of the second granularity; constructing a semantic alignment modulation flow field of the raw material distribution between the first alternative local feature encoding vector of the raw material distribution image of a certain granularity and the second alternative local feature encoding vector of the raw material distribution image of a certain granularity; performing feature alignment on the first alternative local feature encoding vector of the raw material distribution image of a certain granularity and the second alternative local feature encoding vector of the raw material distribution image of a certain granularity based on the semantic alignment modulation flow field of the raw material distribution to obtain an aligned first alternative local feature encoding vector of the raw material distribution image of a certain granularity and an aligned second alternative local feature encoding vector of the raw material distribution image of a certain granularity; performing causal chain loop unit screening on the aligned first alternative local feature encoding vector of the raw material distribution image of a certain granularity and the aligned second alternative local feature encoding vector of the raw material distribution image of a certain granularity to obtain a closed-loop strength factor of the causal chain loop unit of the raw material distribution; and determining whether the first alternative local feature encoding vector of the raw material distribution image of a certain granularity and the second alternative local feature encoding vector of the raw material distribution image of a certain granularity form the causal chain loop unit of the raw material distribution based on the comparison between the closed-loop strength factor of the causal chain loop unit of the raw material distribution and a preset threshold. The above process can be expressed by the formula: ; where represents matrix multiplication. Indicates a transpose operation, Indicates performing convolutional encoding with a 3×3 convolutional kernel, Indicates point convolutional encoding, Indicates a semantic alignment modulation flow field for granular raw material distribution, Indicates The corresponding local feature encoding vector of the first alternative granular raw material distribution image after alignment, Indicates The corresponding local feature encoding vector of the second alternative granular raw material distribution image after alignment, Indicates the inverse hyperbolic cosine function, Indicates the square of the Euclidean norm of a vector, Indicates And The closed-loop strength factor of the causal chain link unit for granular raw material distribution therebetween, Indicates a preset threshold, Indicates And The {first local feature encoding vector of the granular raw material distribution image, second local feature encoding vector of the granular raw material distribution image} feature pair corresponding thereto, Indicates the set of {first local feature encoding vector of the granular raw material distribution image, second local feature encoding vector of the granular raw material distribution image} feature pairs that make up the causal chain link unit for granular raw material distribution.

[0039] It should be understood that the set of local feature encoding vectors of the first-granularity raw material distribution image and the set of local feature encoding vectors of the second-granularity raw material distribution image are respectively from local image feature maps of different sizes and have different semantic background dimensions. If they are directly matched, due to the inconsistent information distribution, it is easy to cause semantic mismatches or offsets. For example, directly matching a feature encoding vector representing the macroscopic distribution trend of large particles with a feature encoding vector representing the microscopic details of small particles may cause the model to have a deviation in the semantic understanding of these features and be unable to accurately judge the relationship between them and the impact on the uniformity of the raw material particle size. Therefore, it is necessary to process the feature pairs through a certain method to avoid the occurrence of such semantic problems. Based on this, the features are precisely dynamically aligned by calculating the semantic alignment modulation flow field of the granular raw material distribution. After achieving feature alignment, the screening algorithm of the causal chain loop unit screening network is used to screen whether the features meet the causal coupling conditions. The core of this screening process is to calculate the closed-loop strength factor of the granular raw material distribution causal chain loop unit between feature pairs to capture those highly relevant and causally related feature pairs. Specifically, those feature pairs with a low closed-loop strength factor of the granular raw material distribution causal chain loop unit are considered noise feature pairs and will be filtered out. This can effectively remove the noise interference in the data and make the model more focused on those truly valuable feature pairs. For example, some accidentally occurring feature pairs that are irrelevant to the uniformity of the raw material particle size will be screened out to improve the quality of the feature data and the analysis efficiency of the model. And those feature pairs with high causality will form the basic structure of the causal chain loop unit. By capturing these highly relevant and causally related feature pairs, the model can better understand the internal mechanism of the raw material particle size distribution. For example, there may be a causal relationship between the shape feature of particles and the dispersibility of particles in the reaction, and the screening network will retain these causally related feature pairs to provide stronger support for the subsequent judgment of the uniformity of the raw material particle size, thereby improving the accuracy and reliability of the model judgment.

[0040] Specifically, the closed-loop strength factor of the granular raw material distribution causal chain loop unit needs to further interact with the respective distribution environment bases of the features on the basis of evolving knowledge bit by bit among the memory feature pairs. That is, the closed-loop strength factor of the granular raw material distribution causal chain loop unit includes a knowledge storage part based on the differential norm of the feature vectors and the respective norm representations of the feature vectors of the environmental interaction part below, in order to cope with the forgetting of causal evolution knowledge and the loss of self-correlation plasticity. Further, on the basis of the positive contribution (as the numerator) of the evolving knowledge of the distribution to the closed-loop strength of the granular raw material distribution causal chain loop and the negative environmental interaction reverse contribution (as the denominator), it is quantified and represented by the geometric correlation causal strength ( Use a function to visually represent the directional similarity of the causal closed-loop strength of the granular raw material distribution, so as to obtain the causal chain link unit closed-loop strength factor of the granular raw material distribution for screening the feature pairs of {the first local feature encoding vector of the granular raw material distribution image, the second local feature encoding vector of the granular raw material distribution image} that are highly relevant and have a causal relationship.

[0041] Specifically, in the embodiment of the present application, the step 253 includes: performing a point-by-point subtraction operation and a point-by-point addition operation on the first local feature encoding vector of the granular raw material distribution image and the second local feature encoding vector of the granular raw material distribution image corresponding to each feature pair in the set of feature pairs of {the first local feature encoding vector of the granular raw material distribution image, the second local feature encoding vector of the granular raw material distribution image} that constitutes the causal chain link unit of the granular raw material distribution to obtain a set of granular raw material distribution image difference local feature vectors and a set of granular raw material distribution image aggregation local feature vectors; multiplying each granular raw material distribution image difference local feature vector in the set of granular raw material distribution image difference local feature vectors by the first weight matrix to obtain a set of granular raw material distribution image difference local modulation feature vectors; multiplying each granular raw material distribution image aggregation local feature vector in the set of granular raw material distribution image aggregation local feature vectors by the second weight matrix to obtain a set of granular raw material distribution image aggregation local modulation feature vectors; performing weighted fusion on each pair of corresponding granular raw material distribution image difference local modulation feature vectors and granular raw material distribution image aggregation local modulation feature vectors in the set of granular raw material distribution image difference local modulation feature vectors and the set of granular raw material distribution image aggregation local modulation feature vectors to obtain the set of heterogeneous granular raw material distribution image semantic-level alignment encoding feature vectors. The above process can be expressed by the formula: ; where represents point-by-point subtraction, represents point-by-point addition, and represent the first weight matrix and the second weight matrix respectively, and represent weighted hyperparameters, represents and the heterogeneous granular raw material distribution image semantic-level alignment encoding feature vector between.

[0042] It should be understood that after the set of feature pairs {local feature encoding vectors of the first-granularity raw material distribution image, local feature encoding vectors of the second-granularity raw material distribution image} is input into the causal link unit semantic dynamic response network, the network, through high-order dynamic interaction modeling, organically integrates the individual semantics represented by the two vectors in each feature pair (such as the granularity size and shape of particles) with the potential causal relationship between them (for example, the particle shape may affect its dispersion in the reaction, and thus affect the granularity uniformity). That is, previous data processing may only consider the semantics of each feature separately, but now through the causal link unit semantic dynamic response network, these semantic information are interrelated within the framework of causal relationships, forming a more meaningful feature representation, that is, the set of semantic-level aligned encoding feature vectors of the heterogeneous-granularity raw material distribution image is generated. Specifically, when the local features of the granularity raw material distribution image are input, the causal link unit semantic dynamic response network can generate a set of dynamic weights according to the feature semantic context, so as to capture the non-linear complex interactions between features. When the semantic of a feature pair {local feature encoding vectors of the first-granularity raw material distribution image, local feature encoding vectors of the second-granularity raw material distribution image} that constitutes a causal link unit of the granularity raw material distribution changes, the network can adjust the output in real time through the dynamic weights, and imitate the response mechanism in the actual causal link in an event-triggered manner. This dynamically enhanced method effectively enriches the relationship modeling ability between the local features of the granularity raw material distribution image, and enables the set of semantic-level aligned encoding feature vectors of the heterogeneous-granularity raw material distribution image to have stronger expression ability at a broader semantic and causal level.

[0043] Finally, the set of semantic-level aligned encoding feature vectors of the heterogeneous-granularity raw material distribution image is subjected to feature aggregation to obtain the semantic-level aligned encoding feature map of the heterogeneous-granularity raw material distribution image. The above process can be expressed by the formula: ; where represents the feature aggregation operation, represents the first, second, th, and th semantic-level aligned encoding feature vectors in the set of semantic-level aligned encoding feature vectors of the heterogeneous-granularity raw material distribution image, represents the semantic-level aligned encoding feature map of the heterogeneous-granularity raw material distribution image.

[0044] It should be understood that the set of semantic-level aligned encoding feature vectors of the heterogeneous granularity raw material distribution images contains rich information from local feature encoding vector pairs of different granularity raw material distribution images. These vectors describe the granularity distribution characteristics of the raw materials from multiple perspectives, such as particle size, shape, distribution density, etc. Through feature aggregation, these scattered multi-dimensional information is integrated into the semantic-level aligned encoding feature map of the heterogeneous granularity raw material distribution images, enabling the feature map to more comprehensively reflect the overall situation of the raw material granularity. For example, the information about large particles and small particles that was originally scattered in different vectors can be presented in a comprehensive manner in the feature map after aggregation, facilitating the subsequent overall judgment and analysis of the uniformity of the raw material granularity.

[0045] In step 26, based on the semantic-level aligned encoding feature map of the heterogeneous granularity raw material distribution images, an identification result is obtained, and the identification result is used to indicate whether the particle uniformity of the preprocessed raw materials meets the preset requirements. Specifically, in the embodiment of the present application, step 26 includes: inputting the semantic-level aligned encoding feature map of the heterogeneous granularity raw material distribution images into a state recognizer based on a classifier to obtain the identification result. More specifically, in the embodiment of the present application, inputting the semantic-level aligned encoding feature map of the heterogeneous granularity raw material distribution images into a state recognizer based on a classifier to obtain the identification result includes: expanding each semantic-level aligned encoding feature matrix in the semantic-level aligned encoding feature map of the heterogeneous granularity raw material distribution images into a one-dimensional feature vector according to row vectors or column vectors and then cascading them to obtain a heterogeneous granularity raw material distribution encoding classification feature vector; using the fully connected layer of the classifier to perform fully connected encoding on the heterogeneous granularity raw material distribution encoding classification feature vector to obtain a heterogeneous granularity raw material distribution fully connected encoding classification feature vector; inputting the heterogeneous granularity raw material distribution fully connected encoding classification feature vector into the Softmax classification function of the classifier to obtain the identification result.

[0046] It should be understood that a classifier is a trained model that has the ability to classify input data. By training the classifier with a large number of labeled sample data (i.e., raw material particle size distribution images with known particle uniformity that meet the preset requirements), the classifier can learn the differences between the characteristics of raw material particle size distributions that meet the requirements and those that do not. By inputting the semantically aligned encoded feature map of the heterogeneous particle size raw material distribution image into the state recognizer based on the classifier, its learned patterns and classification capabilities can be used to more accurately and efficiently classify the new feature map, and then determine the particle uniformity of the raw material. The generated recognition result directly reflects the particle uniformity of the preprocessed raw material, which is helpful for quality control in the silicon carbide production process. Specifically, if the recognition result shows that the particle uniformity of the raw material meets the preset requirements, subsequent production steps such as raw material mixing and carbothermal reduction reaction of silicon carbide can be continued to ensure the quality stability of the final product. If the recognition result shows non-compliance, production personnel can take timely measures, such as readjusting the crushing and screening process parameters and reprocessing the raw material until the particle uniformity meets the requirements, thereby avoiding product quality non-compliance caused by raw material quality problems and reducing production waste and cost losses.

[0047] In summary, step 2 is clearly described. It uses image processing technology based on computer vision to perform multi-size screenshot processing on the particle size distribution image of the preprocessed raw material to obtain the first and second local particle size distribution images, then extracts the particle size distribution characteristics of the first and second local particle size distribution images to obtain the first and second particle size raw material distribution image feature maps, and finally intelligently determines whether the particle uniformity of the preprocessed raw material meets the preset requirements based on the raw material particle size distribution alignment analysis representation of the first and second particle size raw material distribution image feature maps. In this way, by precisely controlling the particle size uniformity of the raw material, the sufficiency of the carbothermal reduction reaction can be significantly improved, thereby improving the crystal quality of silicon carbide, reducing impurity content, and reducing crystal defects.

[0048] In step 3, after mixing the pretreated raw materials, put them into a high-temperature furnace for carbothermal reduction reaction to obtain the reaction product. It should be understood that by fully mixing the pretreated raw materials (waste silicon carbide blocks, high-purity quartz stones, and waste graphite powder), it can ensure that each component is evenly distributed during the reaction, so that silicon atoms and carbon atoms can fully contact and react in the subsequent reaction, avoiding the situation of insufficient local reaction or the formation of impurity phases, and ensuring the consistency and stability of the composition of the final silicon carbide product. The mixing process can adjust the proportions of various raw materials according to the optimal ratio determined by experiments to ensure that the stoichiometric relationship between carbon and the silicon source reaches the optimal state, thereby helping to increase the yield and quality of silicon carbide. Then, the pretreated raw materials after mixing need to be put into a high-temperature furnace for carbothermal reduction reaction, which is a key process for silicon carbide preparation. The purpose is to generate silicon carbide through the chemical reaction between silicon atoms and carbon atoms under high-temperature conditions. It should be noted that the reaction temperature and time of the carbothermal reduction reaction here are the key factors determining the crystal structure and properties of silicon carbide. By adjusting parameters such as the reaction temperature and reaction time, the performance parameters of the final silicon carbide product, such as particle size and crystal structure, can be controlled to meet the requirements of different application fields.

[0049] The following is a detailed description of a specific implementation process of "after mixing the pretreated raw materials, put them into a high-temperature furnace for carbothermal reduction reaction to obtain the reaction product": First of all, in the raw material mixing stage, accurate ratio determination is the basis. According to the chemical principle of silicon carbide generation and the empirical data accumulated from a large number of experiments, strict ratio calculations are carried out for the raw materials of waste silicon carbide, high-purity quartz stone and waste graphite powder. This process requires comprehensive consideration of multiple factors, including the purity of each component in the raw material, the impurity content and their chemical activity in the reaction. Only by determining the precise mixing ratio can it be guaranteed that in the subsequent reaction, the carbon and silicon sources react fully to generate high-quality silicon carbide, avoiding problems such as insufficient reaction or excessive impurity generation due to improper ratio. After determining the ratio, professional mixing equipment is required to achieve uniform mixing of the raw materials. Equipment such as double cone mixers and V-type mixers, with their unique structural design, can drive the raw materials to continue to tumble and strongly convect during operation. In this process, the agglomeration phenomenon between the raw materials will be effectively broken, so that each component can be evenly dispersed in the mixture to provide a good material basis for the subsequent carbon thermal reduction reaction. It should be noted that the time and speed control during the mixing process are crucial. If the mixing time is too short, the raw materials are difficult to mix fully, which will inevitably lead to uneven local components during the reaction and affect the reaction effect; if the mixing time is too long, it will not only reduce production efficiency, but also cause some raw materials to produce fine powder due to excessive friction. These fine powders may bring new problems in the reaction, such as affecting the uniformity of the reaction and the purity of the product. The speed adjustment should also be cautious and needs to be determined in combination with the characteristics of the raw materials and the type of mixing equipment selected to achieve the best mixing effect.

[0050] After the raw materials are mixed, the preparation of the high-temperature furnace begins immediately. The selection of the high-temperature furnace needs to determine the appropriate type and specifications according to the scale of production and the specific requirements for the performance of silicon carbide products. Common resistance furnaces and induction furnaces each have their own advantages. Resistance furnaces perform well in temperature control and can achieve high-precision temperature regulation, which is suitable for production scenarios with demanding reaction temperature requirements; induction furnaces have the characteristics of fast heating speed, which can significantly improve production efficiency. Before being put into use, the high-temperature furnace must be fully and carefully inspected and debugged. The sealing inspection of the furnace body is indispensable, because during the reaction process, if air enters the furnace, it will change the reaction atmosphere, causing the raw materials to undergo oxidation reaction at high temperature, which will seriously affect the product quality. At the same time, the heating elements should be checked one by one to ensure that they can work normally and provide a stable and reliable high-temperature environment for the reaction. The debugging of the temperature control system should not be ignored, and key parameters such as the heating rate, insulation time and target reaction temperature need to be accurately set. In addition, in order to create an oxygen-free reaction environment, protective gases such as argon need to be prepared. Before the reaction starts, a protective gas is introduced into the furnace to completely exhaust the air in the furnace to create ideal conditions for the subsequent carbon thermal reduction reaction.

[0051] After all preparations are made, it enters the carbothermal reduction reaction stage. The uniformly mixed raw materials are put into a high-temperature furnace, and the furnace door is closed to ensure good sealing inside the furnace. Subsequently, the high-temperature furnace is started according to the pre-set heating program. Controlling the heating rate is crucial and it should not be too fast because too rapid heating will cause the raw materials to react too violently instantaneously, making it difficult to control effectively, and may also trigger sudden changes in the furnace pressure, posing safety hazards. Usually, according to the characteristics of the raw materials and the specific requirements of the reaction, the heating rate is controlled within a reasonable range, such as rising 5 - 10 °C per minute. When the temperature inside the furnace reaches the specific temperature required for the carbothermal reduction reaction, it enters the heat preservation stage. During heat preservation, silicon atoms and carbon atoms undergo chemical reactions under the action of high temperature and gradually form silicon carbide. The length of the heat preservation time has a profound impact on the crystal structure and properties of silicon carbide and must be precisely controlled according to the actual situation. The time may vary from several hours to more than ten hours. Throughout the reaction process, key parameters such as the temperature and pressure inside the furnace should be closely monitored at all times to ensure the reaction proceeds smoothly under stable conditions. Once abnormal parameters are found, such as too high temperature or too high pressure, corresponding adjustment measures should be taken immediately to ensure the normal progress of the reaction.

[0052] After the reaction ends, the product treatment link cannot be underestimated either. After the reaction is completed, the product cannot be taken out immediately but has to wait for the high-temperature furnace to cool down to a suitable temperature. If the product is taken out eagerly at a high temperature, due to the sharp temperature change, the product is extremely likely to have cracks or other defects, seriously affecting the product quality. After the furnace temperature cools down to an appropriate level, carefully take out the reaction product and conduct a preliminary visual inspection. By observing the appearance form, color and other characteristics of the product, it can be preliminarily judged whether the reaction proceeds normally. Then, transfer the product to the subsequent treatment process. It should be noted especially that during the transfer of the product, great attention should be paid to avoiding contamination or secondary damage to the product to ensure its quality is not affected.

[0053] In step 4, the reaction product is crushed and screened to obtain pure silicon carbide blocks. It should be understood that after the carbothermal reduction reaction is completed in the high-temperature furnace, some raw materials may not be completely converted into silicon carbide, and these residues may include unreacted quartzite, graphite or other impurities. Through crushing and screening, these impurities can be effectively removed to ensure the purity of the final product. That is, the pure silicon carbide blocks obtained after crushing and screening treatment have higher purity, improving the quality of the final silicon carbide product. In addition, an effective crushing and screening process can also classify and utilize silicon carbide with different particle sizes screened out to achieve the maximum utilization of resources and reduce waste.

[0054] In summary, the production method of silicon carbide based on the embodiments of the present application is elucidated. First, raw materials including waste silicon carbide blocks, high-purity quartzite, and waste graphite powder are prepared. Then, these raw materials are crushed and screened to obtain pre-treated raw materials suitable for subsequent processing. Next, the pre-treated mixed raw materials are placed in a high-temperature furnace for carbothermal reduction reaction to generate reaction products. Finally, the reaction products are further crushed and screened to obtain pure silicon carbide blocks. In this way, by recycling waste silicon carbide blocks and waste graphite powder, the effective utilization of resources can be achieved and environmental pollution can be reduced.

Claims

1. A method for producing silicon carbide, characterized in that: include: Step 1: preparing raw materials, the raw materials comprising waste silicon carbide blocks, high-purity quartz stones and waste graphite powder; Step 2: crushing and sieving the waste silicon carbide block, the high-purity quartz stone and the waste graphite powder to obtain pre-treated raw materials; Step 3: After mixing the pretreated raw materials, place them in a high-temperature furnace for carbon thermal reduction reaction to obtain a reaction product; Step 4: crushing and screening the reaction product to obtain pure silicon carbide blocks.

2. The method for producing silicon carbide according to claim 1, characterized in that: The step 2: crushing and screening the waste silicon carbide block, the high-purity quartz stone and the waste graphite powder to obtain pre-treated raw materials, comprising: Acquiring a particle size distribution image of the pretreated raw material captured by a camera; Cutting out a first local particle size distribution image having a first size from the particle size distribution image; Cutting out a second local particle size distribution image having a second size from the particle size distribution image, wherein the second size is smaller than the first size, and the first local particle size distribution image and the second local particle size distribution image have the same shape; Extracting particle size distribution features from the first local particle size distribution image and the second local particle size distribution image to obtain a first particle size raw material distribution image feature map and a second particle size raw material distribution image feature map; Performing a raw material particle size distribution alignment analysis with strong causal connection on the first particle size raw material distribution image feature map and the second particle size raw material distribution image feature map to obtain a semantic level alignment coding feature map of different particle size raw material distribution images; Based on the semantic-level alignment coding feature map of the distribution image of the raw materials with different particle sizes, a recognition result is obtained, and the recognition result is used to indicate whether the particle uniformity of the raw materials after pretreatment meets the preset requirements.

3. The method for producing silicon carbide according to claim 2, characterized in that: The first local particle size distribution image and the second local particle size distribution image are subjected to particle size distribution feature extraction to obtain a first particle size raw material distribution image feature map and a second particle size raw material distribution image feature map, including: inputting the first local particle size distribution image and the second local particle size distribution image into a particle size distribution feature extractor respectively to obtain the first particle size raw material distribution image feature map and the second particle size raw material distribution image feature map.

4. The method for producing silicon carbide according to claim 3, characterized in that: The particle size distribution feature extractor is a particle size distribution feature extractor based on a hollow pyramid model.

5. The method for producing silicon carbide according to claim 4, characterized in that: Performing a raw material particle size distribution alignment analysis with strong causal connection on the first particle size raw material distribution image feature map and the second particle size raw material distribution image feature map to obtain a semantic level alignment coding feature map of different particle size raw material distribution images, including: Decoupling the first particle size raw material distribution image feature map and the second particle size raw material distribution image feature map along their channel dimensions to obtain a set of first particle size raw material distribution image local feature coding vectors and a set of second particle size raw material distribution image local feature coding vectors; Inputting the set of local feature coding vectors of the first particle size raw material distribution image and the set of local feature coding vectors of the second particle size raw material distribution image into a causal chain unit screening network to obtain a set of feature pairs of {first particle size raw material distribution image local feature coding vector, second particle size raw material distribution image local feature coding vector} constituting a particle size raw material distribution causal chain unit; Inputting the set of feature pairs of {first particle size raw material distribution image local feature coding vector, second particle size raw material distribution image local feature coding vector} constituting the particle size raw material distribution causal chain unit into the causal chain unit semantic dynamic response network to obtain a set of semantic level aligned coding feature vectors of different particle size raw material distribution images; The set of semantic-level aligned coding feature vectors of the distribution images of raw materials with different particle sizes is subjected to feature aggregation to obtain a semantic-level aligned coding feature map of the distribution images of raw materials with different particle sizes.

6. The method for producing silicon carbide according to claim 5, characterized in that: Inputting the set of local feature coding vectors of the first particle size raw material distribution image and the set of local feature coding vectors of the second particle size raw material distribution image into the causal chain unit screening network to obtain a set of feature pairs of {first particle size raw material distribution image local feature coding vector, second particle size raw material distribution image local feature coding vector} constituting the particle size raw material distribution causal chain unit, including: Extracting a first candidate particle size raw material distribution image local feature coding vector from the set of the first particle size raw material distribution image local feature coding vectors; Extracting a second candidate particle size raw material distribution image local feature coding vector from the set of the second particle size raw material distribution image local feature coding vectors; Constructing a particle size raw material distribution semantic alignment modulation flow field between the first candidate particle size raw material distribution image local feature coding vector and the second candidate particle size raw material distribution image local feature coding vector; Based on the particle size raw material distribution semantic alignment modulation flow field, feature alignment is performed on the first candidate particle size raw material distribution image local feature coding vector and the second candidate particle size raw material distribution image local feature coding vector to obtain an aligned first candidate particle size raw material distribution image local feature coding vector and an aligned second candidate particle size raw material distribution image local feature coding vector; Performing causal chain unit screening on the aligned local feature coding vector of the first candidate particle size raw material distribution image and the aligned local feature coding vector of the second candidate particle size raw material distribution image to obtain a particle size raw material distribution causal chain unit closed loop strength factor; Based on the comparison between the closed-loop strength factor of the particle size raw material distribution causal chain unit and a preset threshold, it is determined whether the first candidate particle size raw material distribution image local feature coding vector and the second candidate particle size raw material distribution image local feature coding vector constitute the particle size raw material distribution causal chain unit.

7. The method for producing silicon carbide according to claim 6, characterized in that: Inputting the set of feature pairs of {first particle size raw material distribution image local feature coding vector, second particle size raw material distribution image local feature coding vector} constituting the particle size raw material distribution causal chain unit into the causal chain unit semantic dynamic response network to obtain a set of semantic level aligned coding feature vectors of different particle size raw material distribution images, including: Performing position point subtraction operations and position point addition operations on the first particle size raw material distribution image local feature coding vector and the second particle size raw material distribution image local feature coding vector corresponding to each feature pair in the set of {first particle size raw material distribution image local feature coding vector, second particle size raw material distribution image local feature coding vector} constituting the particle size raw material distribution causal chain unit, respectively, to obtain a set of particle size raw material distribution image difference local feature vectors and a set of particle size raw material distribution image aggregation local feature vectors; Respectively multiplying each particle size raw material distribution image difference local feature vector in the particle size raw material distribution image difference local feature vector set with a first weight matrix to obtain a particle size raw material distribution image difference local modulation feature vector set; Multiplying each particle size raw material distribution image aggregation local feature vector in the particle size raw material distribution image aggregation local feature vector set by the second weight matrix to obtain a particle size raw material distribution image aggregation local modulation feature vector set; The set of local modulation feature vectors of particle size raw material distribution difference and the set of local modulation feature vectors of particle size raw material distribution aggregation are weighted fused to obtain the set of semantic-level aligned coding feature vectors of heterogeneous particle size raw material distribution images.

8. The method for producing silicon carbide according to claim 7, characterized in that: Based on the semantic-level aligned coding feature map of the distribution image of raw materials with different particle sizes, a recognition result is obtained, including: inputting the semantic-level aligned coding feature map of the distribution image of raw materials with different particle sizes into a state recognizer based on a classifier to obtain the recognition result.

9. The method for producing silicon carbide according to claim 8, characterized in that: Inputting the semantic level alignment coding feature map of the distribution image of the raw materials with different particle sizes into the state recognizer based on the classifier to obtain the recognition result, including: Expanding each semantic-level alignment coding feature matrix of the raw material distribution image of different particle sizes in the semantic-level alignment coding feature map of the raw material distribution image of different particle sizes into a one-dimensional feature vector according to a row vector or a column vector, and then cascading them to obtain a distribution coding classification feature vector of the raw material distribution of different particle sizes; Using the fully connected layer of the classifier to perform fully connected encoding on the distribution encoding classification feature vector of the raw materials with different particle sizes to obtain the distribution fully connected encoding classification feature vector of the raw materials with different particle sizes; The fully connected encoded classification feature vector of the different particle size raw material distribution is input into the Softmax classification function of the classifier to obtain the recognition result.

Images