Method and system for controlling CVD (Chemical Vapor Deposition) growth of silicon carbide particles based on multi-mode perception

Abstract

The invention discloses a silicon carbide particle CVD growth control method and system based on multi-modal sensing, and relates to the technical field of intelligent manufacturing of new materials. The multi-modal data comprises a particle morphology image, process parameters and crystal form characterization data; identifying the particle morphology image to obtain a real-time particle size; the particle size deviation is calculated based on the real-time particle size, a crystal form optimization mark is calculated based on the crystal form characterization data, if the particle size deviation is larger than a preset particle size deviation threshold value or the crystal form optimization mark is true, an improved AI decision engine is triggered to make a compensation decision, and the optimal adjusting amount of the technological parameters is obtained and executed; through multi-source real-time feedback, an improved AI decision engine and multi-objective optimization, the problems of over-wide particle size distribution and high energy consumption of a traditional process are solved.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing technology for new materials, and in particular to a method and system for controlling the CVD growth of silicon carbide particles based on multimodal sensing. Background Technology

[0002] The statements in this section are merely background information relating to this disclosure and do not necessarily constitute prior art.

[0003] Silicon carbide (SiC) has become a core material for third-generation semiconductors due to its high thermal conductivity, wide bandgap, and high-temperature resistance. Currently, mainstream SiC powder mass production relies on chemical vapor deposition (CVD), which uses methyltrichlorosilane (MTS) for pyrolysis deposition in a high-temperature reactor to achieve high-purity particle preparation. Existing CVD processes generally use preset temperature, pressure, and flow parameters, making them unable to respond to micro-disturbances within the furnace, such as temperature fluctuations (due to heating element aging or airflow erosion causing axial temperature differences) and pressure drift (due to deposition byproducts clogging exhaust pipes). These disturbances can directly trigger differences in nucleation rates, leading to particle agglomeration and crystal distortion.

[0004] Furthermore, there is a particle size dispersion problem caused by lag control, and traditional PID temperature regulation has a significant delay. For example, there is a certain delay in thermocouple detection (S-type thermocouple response time), PID calculation and power adjustment, and temperature conduction to the reaction zone. The cumulative lag time is long, resulting in high particle size deviation.

[0005] Current CVD processes also present challenges due to multi-parameter coupling, with strong coupling effects between temperature (T), pressure (P), and gas flow rate ratio (R). For example, in the low-temperature region (T<1200℃), particle size is sensitive to temperature but insensitive to pressure changes; in the critical deposition region (T≥1300℃), increasing the MTS flow rate can increase particle size, but decrease temperature sensitivity. Current technologies lack dynamic decoupling capabilities, making synergistic optimization difficult.

[0006] Offline testing also suffers from time-sensitivity issues. The industry generally uses sampling offline testing (such as laser particle size analyzers), which requires stopping the furnace to cool down for sampling. Sampling is done at intervals, which is time-consuming and has a delay in feedback. This causes real-time control to fail completely, resulting in a consistently low pass rate and becoming a bottleneck in the manufacturing of high-precision powders. Summary of the Invention

[0007] To overcome the shortcomings of the prior art, this invention provides a method and system for controlling the CVD growth of silicon carbide particles based on multimodal sensing. Through multi-source real-time feedback, an improved AI decision engine, and multi-objective optimization, it solves the problems of excessively wide particle size distribution and high energy consumption in traditional processes.

[0008] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: In a first aspect, the present invention provides a method for controlling the CVD growth of silicon carbide particles based on multimodal sensing, comprising: Simultaneously acquire multimodal data inside the reaction chamber; the multimodal data includes particle morphology images, process parameters, and crystal form characterization data; The particle morphology image is identified to obtain the real-time particle size; The particle size deviation is calculated based on real-time particle size, and the crystal form optimization flag is calculated based on crystal form characterization data. If the particle size deviation is greater than the preset particle size deviation threshold or the crystal form optimization flag is true, the improved AI decision engine is triggered to make a compensation decision, obtain the optimal adjustment amount of the process parameters, and execute it. The improved AI decision engine employs a hybrid reasoning mechanism that combines deep reinforcement learning and fuzzy logic to construct a multi-objective comprehensive cost function, with minimizing the multi-objective comprehensive cost function as the optimization objective.

[0009] A further technical solution involves embedding a high-temperature resistant imaging probe on the inner wall of the reaction chamber to collect particle morphology images, embedding a Raman spectroscopy probe to collect crystal form characterization data, and deploying a distributed sensor array on the top of the reaction chamber to collect process parameters.

[0010] A further technical solution, the specific steps of the improved edge recognition algorithm are as follows: The effective area of ​​the reaction zone is obtained by cropping the field of view of the particle morphology image, and the image is converted into a grayscale image; Adaptive median filtering is applied to the grayscale image to denoise it, and the denoised image is output. Perform contrast-limited adaptive histogram equalization on the denoised image to output an enhanced image; Calculate the horizontal and vertical gradients on the enhanced image to obtain the gradient magnitude and gradient direction; perform non-maximum suppression on the gradient magnitude along the gradient direction to obtain a thinned edge response map; Adaptive double-threshold segmentation is applied to the refined edge response map to output a binary edge map; Morphological processing and connected component filtering are performed on the binary edge graph to form a cleaned edge graph; Extract a set of closed contours from the cleanup edge map, calculate the particle area of ​​each contour in the set, obtain the equivalent diameter based on the particle area, and then statistically analyze the particle size distribution and real-time particle size.

[0011] A further technical solution is that the hybrid reasoning mechanism integrating deep reinforcement learning and fuzzy logic is specifically as follows: Based on particle size deviation, particle size standard deviation, crystal form characterization data, process parameters, energy consumption, deposition rate, and structural state vectors of the production stage; The sensitivity weight vector for the compensation channel is obtained by inferring the state vector based on the dynamic sensitivity weight model. The multi-objective optimization weights are dynamically determined based on the production stage, and the particle size deviation, energy consumption and deposition rate are dynamically weighed in the multi-objective comprehensive cost function. Generate prior constraint or compensation channel preference vectors using fuzzy logic reasoning; Actions are selected under the compensation channel preference vector or prior constraints, and the optimal compensation strategy is generated through deep reinforcement learning.

[0012] A further technical solution is that the dynamic sensitivity weight model is constructed based on a gradient boosting decision tree to obtain a sensitivity weight vector for the compensation channel, expressed as:

[0013] in, For the sensitivity weight vector, , , The relative sensitivity of temperature, pressure, and flow ratio to particle size deviation correction is characterized, respectively.

[0014] A further technical solution defines the action as follows:

[0015] in, For action, For temperature regulation, For pressure regulation, This is the flow rate adjustment amount.

[0016] A further technical solution is that the multi-objective integrated cost function is expressed as:

[0017] in, For particle size deviation, For energy consumption, For the deposition rate, For particle size standard deviation, , , To follow the production stage Dynamically changing weights , The weights of the particle size dispersion and crystal form penalty terms are given. The crystal form threshold. As a crystal form index, This indicates a penalty for failing to meet crystal form standards.

[0018] Secondly, the present invention provides a silicon carbide particle CVD growth control system based on multimodal sensing, comprising: The multi-source data acquisition module is configured to: synchronously acquire multimodal data inside the reaction chamber; the multimodal data includes particle morphology images, process parameters, and crystal form characterization data; The particle size recognition module is configured to: recognize the particle morphology image to obtain the real-time particle size; The compensation decision module is configured to: calculate the particle size deviation based on real-time particle size, calculate the crystal form optimization flag based on crystal form characterization data, and if the particle size deviation is greater than the preset particle size deviation threshold or the crystal form optimization flag is true, trigger the improved AI decision engine to make a compensation decision, obtain the optimal adjustment amount of the process parameters and execute it. The improved AI decision engine employs a hybrid reasoning mechanism that combines deep reinforcement learning and fuzzy logic to construct a multi-objective comprehensive cost function, with minimizing the multi-objective comprehensive cost function as the optimization objective.

[0019] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the multimodal sensing-based silicon carbide particle CVD growth control method described in the first aspect.

[0020] Fourthly, the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in the multimodal sensing-based silicon carbide particle CVD growth control method described in the first aspect.

[0021] The above one or more technical solutions have the following beneficial effects: This invention achieves coordinated optimization control of temperature, pressure, and gas flow rate through multi-source information fusion and dynamic sensitivity weighting model; the AI ​​decision engine adopts a hybrid reasoning mechanism that combines deep reinforcement learning and fuzzy logic, and introduces a multi-objective optimization function to dynamically balance energy consumption and deposition rate while ensuring particle size accuracy.

[0022] This invention integrates a multimodal sensing unit consisting of a high-temperature imaging probe, a laser scattering sensor, and a Raman spectroscopy probe, and combines it with an improved edge recognition algorithm. This enables the high-precision acquisition of particle morphology, particle size, submicron particle concentration, and microcrystalline structure information in harsh high-temperature reaction environments, providing a rich and reliable data foundation for control decisions.

[0023] The method of this invention reduces the standard deviation of particle size, narrows the ratio of particle size distribution parameters, improves the yield, and reduces energy consumption per unit output, thus solving the problem of excessively wide particle size distribution caused by response lag and parameter coupling in traditional CVD processes. Attached Figure Description

[0024] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0025] Figure 1 This is a flowchart of a silicon carbide particle CVD growth control method based on multimodal sensing, according to an embodiment of the present invention. Detailed Implementation

[0026] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0027] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0028] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0029] Example 1 like Figure 1 As shown, this embodiment discloses a method for controlling the CVD growth of silicon carbide particles based on multimodal sensing. The method includes the following steps: S1: Simultaneously acquire multimodal data inside the reaction chamber; the multimodal data includes particle morphology images, process parameters, and crystal form characterization data; In this embodiment, the main body of the system is a double-layer water-cooled reactor, with a multi-channel gas path system integrated inside the furnace. The gas path system connects the raw material premixing chamber, the reaction chamber, and the waste gas recovery device, forming a closed reaction environment. The system precisely controls the flow ratio of methyltrichlorosilane to argon through a three-way proportional valve, and the variable frequency heating power supply regulates the temperature gradient of the reaction chamber.

[0030] The inner wall of the reaction chamber is embedded with a multimodal sensing unit, which integrates a high-temperature imaging probe, a laser scattering sensor, and a Raman spectroscopy probe. A distributed sensor array is deployed on the top of the reaction chamber, including 6 sets of S-type thermocouples and 3 piezoresistive pressure sensors, forming a 10Hz high-frequency data acquisition network that covers the axial and radial areas of the reactor.

[0031] The high-temperature resistant imaging probe has an tilt angle of 30℃-60℃ and captures particle morphology images at a rate of 10 frames per second. The improved edge recognition algorithm enhances and recognizes the images, and outputs real-time particle size and particle size distribution.

[0032] Laser scattering sensors are used to detect submicron-sized particles (0.1-1 μm) within the blind zone of an imaging probe. It can simultaneously output submicron-level particle concentration, making up for the shortcomings of single imaging technology.

[0033] Raman spectroscopy probes are used for real-time monitoring of particle crystallization quality and / Crystal form ratio, if When the proportion falls below a set threshold (e.g., 85%), a crystal form optimization flag is triggered, providing crystal form dimension feedback for process optimization.

[0034] S2: Identify the particle morphology image to obtain the real-time particle size; In this embodiment, the imaging probe acquires particle morphology images (raw frame images, which can be RGB) at a sampling frequency of approximately 10 fps. The imaging probe is equipped with a nitrogen purging device to prevent mirror contamination.

[0035] The specific steps of the improved edge recognition algorithm are as follows: (1) The field of view of the particle morphology image is cropped to obtain the effective region of reaction (ROI), and the RGB image is converted into a grayscale image.

[0036] (2) Perform adaptive median filtering on the grayscale image to denoise it, preserving edges and removing impulse noise, and output the denoised image. For each pixel in the grayscale image... With window (by minimum window) Gradually increase to the maximum window size ), calculate the minimum value within the window Maximum value Median and current pixel value It achieves adaptive suppression of impulse noise through two-level judgment.

[0037] The judgment rule is: Level 1 (Determine if the median is noise): ,

[0038] like and If it succeeds, proceed to the second-level judgment; otherwise, increase the window size. until .

[0039] Level 2 (determines whether the current pixel is noise): ,

[0040] like and Output Otherwise output .

[0041] (3) Perform contrast-limited adaptive histogram equalization (CLAHE) on the denoised image and output the enhanced image. Divide the denoised image into several sub-blocks and perform histogram equalization on the sub-blocks; at the same time, set a clipping threshold to limit the peak value of the histogram and suppress the amplification of noise; finally, perform inter-block interpolation for smooth transition.

[0042] (4) Calculate the horizontal gradient on the enhanced image with vertical gradient This yields the gradient magnitude and gradient direction. The formula is expressed as:

[0043]

[0044] in, For pixels gradient, For pixels The horizontal gradient, For pixels The vertical gradient, The gradient direction for pixel XY represents the direction in which the image brightness (or intensity) changes most rapidly at that point. Outputs a gradient magnitude map and a gradient direction map.

[0045] (5) Non-maximum suppression of gradient magnitude is performed along the gradient direction, and only the local maximum gradient response is retained to obtain the refined edge response map.

[0046] (6) Adaptive dual-threshold segmentation is applied to the refined edge response map to output a binary edge map (obtained by hysteresis connection of strong and weak edges). A high threshold is constructed using the statistics of the refined edge response map (such as the median or quantiles) as the threshold benchmark. and low threshold Furthermore, hysteresis connections are used to preserve weak edges that are connected to strong edges, thereby improving edge continuity and reducing false edges.

[0047] If the median is used as the threshold, it is expressed as:

[0048]

[0049]

[0050] in, The median, To refine the edge response map, , This is an adjustable coefficient.

[0051] (7) Perform morphological processing and connected component filtering on the binary edge map to eliminate isolated noise points, repair edge breaks, and generate a cleaned edge map. Common operations include opening, closing, hole filling, and area thresholding.

[0052] (8) Extract the set of closed contours from the cleaned-up edge map (Effective particle profile set). Adhesive particles are separated by distance exchange and watershed or concave point segmentation; and effective particle profiles are selected based on shape constraints (such as area, roundness, aspect ratio, etc.).

[0053] (9) For the set of closed contours Each outline in Calculate the particle area and obtain the equivalent diameter based on the area. Then, the particle size distribution and real-time particle size are statistically analyzed. The real-time particle size and particle size distribution parameters are used to compare with the target particle size. Compare and trigger subsequent closed-loop control.

[0054] The formula for the equivalent diameter is expressed as:

[0055] in, For outline The corresponding particle area (which can be converted to physical units by multiplying the pixel area by the calibration ratio).

[0056] Real-time particle size definition ratio: If we take the average:

[0057] in, This is the real-time particle size; if the median diameter is taken (in... express) Particle size distribution parameters (such as , , wait); This represents the number of outlines.

[0058] S3: Calculate particle size deviation based on real-time particle size and crystal form optimization flag based on crystal form characterization data. If the particle size deviation is greater than the preset particle size deviation threshold or the crystal form optimization flag is true, the improved AI decision engine is triggered to make a compensation decision, obtain the optimal adjustment amount of process parameters and execute it.

[0059] In this embodiment, the system is equipped with a central control cabinet, which has a built-in improved AI decision engine that receives multimodal data, real-time particle size, and real-time data from the sensor array.

[0060] The improved AI decision engine adopts a hybrid reasoning mechanism that integrates deep reinforcement learning and fuzzy logic to construct a state-action value function. The network structure includes three fully connected hidden layers (128, 64, and 32 neurons), with the optimization objective being to minimize the particle size standard deviation and the energy consumption weighted sum.

[0061] When particle size deviation is detected, the improved AI decision engine introduces a multi-objective optimization function, dynamically weighting particle size accuracy, energy consumption, and deposition rate to generate the optimal compensation strategy.

[0062] Multi-objective optimization function:

[0063] in, , , These are weighting coefficients, dynamically adjusted according to the production stage; Represents the absolute value of dimensional deviation. Indicates energy consumption. This indicates the material deposition rate.

[0064] In some implementations, the improved AI decision engine dynamically selects the dominant compensation parameters based on the operating condition range and calls the dynamic sensitivity weight model to calculate the specific compensation amount: When T < 1200℃ and P ≥ 1.03MPa, the temperature compensation mode is activated, and the heating power adjustment is ΔD × 0.35% / ℃; When T≥1200℃ and P<1.03MPa, the pressure compensation mode is activated, and the exhaust valve opening adjustment is ΔD×0.25% / kPa; When T≥1200℃ and P≥1.03MPa, the flow compensation mode is activated, and the MTS / Ar proportional adjustment is ΔD×0.45% / (L / min).

[0065] The specific processing steps of the improved AI decision engine are as follows: (1) Acquire multi-source data and align them in time to obtain the set of observations.

[0066] Get the real-time particle size of the output To obtain the process parameters, i.e., temperature, output from the sensor array. ,pressure Flow ratio Acquire crystal form characterization data The acquired data are aligned according to a unified timestamp to obtain the set of observations. .

[0067] (2) Calculate each feature and determine the deviation.

[0068] Calculation of particle size deviation based on observation set Particle size standard deviation and generate crystal form markers. .

[0069] The particle size deviation and standard deviation are calculated and expressed as follows:

[0070]

[0071] in, For particle size deviation, for Real-time particle size at any given moment for Target particle size at time, For particle size standard deviation, For the number of particles, For the first Each particle Target particle size at time, for The average target particle size at time t.

[0072] Crystal form optimization markers (example):

[0073] in, This is an indicator function; it takes the value 1 if the condition is true, and 0 otherwise. Crystal form index ( Proportion).

[0074] If the triggering condition is met, proceed to the compensation decision step. The triggering condition can be at least one of the following:

[0075] or

[0076] in, Set a preset particle size deviation threshold (e.g., 0.1). ).

[0077] (3) Define state variables.

[0078] The features used for decision-making at the current moment are used to form a state vector. , is represented as:

[0079] in, Energy consumption index (calculated based on the power of production equipment, which can be energy consumption per unit time or energy consumption per unit output). For the deposition rate, For the production stage (initial nucleation stage, stable growth stage, and later optimization stage), dynamic weight switching is used.

[0080] (4) Calculate the dynamic sensitivity weight.

[0081] A dynamic sensitivity weight model is constructed based on a gradient boosting decision tree and trained using historical accumulated data.

[0082] Using a gradient boosting decision tree (GBDT) model trained on historical sedimentary data, the state vector is... Inference is performed to obtain the sensitivity weight vector to the compensation channel. , is represented as:

[0083] in, , , These parameters represent the relative sensitivity of temperature, pressure, and flow ratio to particle size deviation correction. The parameter with the highest relative sensitivity is the dominant compensation parameter, which is the fundamental variable for performing the calculation.

[0084] In some implementations, the dynamic sensitivity weighting model assigns coefficients according to the following rules: Low temperature domain ( ): =0.8, =0.1, =0.1; Low-pressure area ( ): =0.2, =0.7, =0.1; Critical sedimentary zone: =0.2, =0.3, =0.5.

[0085] The dynamic sensitivity weight model updates the weight coefficients online at fixed intervals (e.g., once every 24 hours) to adapt to process drift.

[0086] (5) Set multi-objective weights.

[0087] Multi-objective optimization weights are dynamically determined based on the production stage. , , It is used to make dynamic trade-offs between particle size accuracy, energy consumption and deposition rate.

[0088] (6) Generate the optimal compensation strategy through deep reinforcement learning.

[0089] Construct a Q-network with a state-action value function, whose network structure includes three fully connected hidden layers, and select actions based on the preference vector or constraints output by fuzzy logic. In order to minimize the overall cost of multiple objectives or maximize the equivalent reward.

[0090] Define Action (Compensation Amount), the action is defined as the adjustment vector for each compensation channel, expressed as:

[0091] in, For temperature regulation, For pressure regulation, This is the flow rate adjustment amount.

[0092] To ensure compatibility with linear compensation and improve feasibility, the action can be set as a scaling factor of the basic compensation, expressed as: , ,

[0093] in, , , The action coefficients are selected by the Q-network. , , These are the basic compensation values ​​for temperature regulation, pressure regulation, and flow ratio regulation, which can be adjusted based on particle size deviation. It is calculated using a preset proportional coefficient.

[0094] Multi-objective integrated cost function Represented as:

[0095] in, , , As the stage Dynamically changing weights , The weights of the particle size dispersion and crystal form penalty terms are given. The crystal form threshold. This indicates a penalty for failing to meet crystal form standards.

[0096] The reward function is defined as the negative of the cost function:

[0097] in, This is the reward value.

[0098] The Q network is represented as:

[0099] in, For online network parameters, It is a constant. As a discount factor, In time step On top of that, add a constant k to execute the action. The instant reward received afterward.

[0100] The objective value of the Q network is defined as: ) in, In time step Execute action The instant reward obtained afterward for The state at any given moment, These are the target network parameters.

[0101] The loss function is defined as:

[0102] in, The loss value measures the difference between the target Q-value and the current Q-value predicted by the network. By minimizing this loss, the online network's predictions gradually approach the target value, thus learning the optimal Q-function.

[0103] (7) Perform optimal compensation and perform closed-loop verification.

[0104] Actions The parameters are converted into adjustments for temperature, pressure, and flow ratio and executed. After compensation, a new round of particle size detection is initiated after a set time (e.g., 60 seconds). Particle size and submicron particle distribution are re-measured, and crystal quality is verified using a Raman spectroscopy probe. Specifically, data is collected again after the preset verification window. ,get If the particle size is still out of tolerance or the crystal form does not meet the standard, secondary compensation is initiated and recorded in the process database for model updates.

[0105] The system includes a compensation execution unit, which contains three sets of linkage mechanisms, including a three-way proportional valve and a variable frequency heating power supply. The gas flow rate regulation accuracy reaches ±0.1L / min, the temperature control response time is <5 seconds, and it executes the instructions issued by the AI ​​decision engine.

[0106] The process database stores four types of core mapping relationships: temperature-particle size sensitivity curves, pressure-particle size response matrices, historical compensation efficiency tables, and historical records of multi-objective optimization weights.

[0107] The closed-loop control process is as follows: multimodal data acquisition → data fusion and deviation analysis → operating condition zoning → multi-objective weight calculation → process parameter compensation → result verification (including crystal form verification), realizing the coordinated optimization of multiple indicators such as particle size, crystal form and energy consumption.

[0108] Example 2 This embodiment discloses a silicon carbide particle CVD growth control system based on multimodal sensing, including: The multi-source data acquisition module is configured to: synchronously acquire multimodal data inside the reaction chamber; the multimodal data includes particle morphology images, process parameters, and crystal form characterization data; The particle size recognition module is configured to: recognize the particle morphology image to obtain the real-time particle size; The compensation decision module is configured to: calculate the particle size deviation based on real-time particle size, calculate the crystal form optimization flag based on crystal form characterization data, and if the particle size deviation is greater than the preset particle size deviation threshold or the crystal form optimization flag is true, trigger the improved AI decision engine to make a compensation decision, obtain the optimal adjustment amount of the process parameters and execute it. The improved AI decision engine employs a hybrid reasoning mechanism that combines deep reinforcement learning and fuzzy logic to construct a multi-objective comprehensive cost function, with minimizing the multi-objective comprehensive cost function as the optimization objective.

[0109] Example 3 The purpose of this embodiment is to provide a computing device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method of Embodiment 1.

[0110] Example 4 The purpose of this embodiment is to provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the steps of the method of Embodiment 1.

[0111] The steps and methods involved in the apparatuses of Embodiments 3 and 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0112] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0113] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0114] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for controlling the CVD growth of silicon carbide particles based on multimodal sensing, characterized in that, include: Simultaneously acquire multimodal data inside the reaction chamber; the multimodal data includes particle morphology images, process parameters, and crystal form characterization data; The particle morphology image is identified to obtain the real-time particle size; The particle size deviation is calculated based on real-time particle size, and the crystal form optimization flag is calculated based on crystal form characterization data. If the particle size deviation is greater than the preset particle size deviation threshold or the crystal form optimization flag is true, the improved AI decision engine is triggered to make a compensation decision, obtain the optimal adjustment amount of the process parameters, and execute it. The improved AI decision engine employs a hybrid reasoning mechanism that combines deep reinforcement learning and fuzzy logic to construct a multi-objective comprehensive cost function, with minimizing the multi-objective comprehensive cost function as the optimization objective.

2. The method for controlling the CVD growth of silicon carbide particles based on multimodal sensing as described in claim 1, characterized in that, A high-temperature resistant imaging probe is embedded in the inner wall of the reaction chamber to collect particle morphology images, a Raman spectroscopy probe is embedded to collect crystal form characterization data, and a distributed sensor array is deployed at the top of the reaction chamber to collect process parameters.

3. The method for controlling the CVD growth of silicon carbide particles based on multimodal sensing as described in claim 1, characterized in that, The specific steps of the improved edge recognition algorithm are as follows: The effective area of ​​the reaction zone is obtained by cropping the field of view of the particle morphology image, and the image is converted into a grayscale image; Adaptive median filtering is applied to the grayscale image to denoise it, and the denoised image is output. Perform contrast-limited adaptive histogram equalization on the denoised image to output an enhanced image; Calculate the horizontal and vertical gradients on the enhanced image to obtain the gradient magnitude and gradient direction; Non-maximum suppression of the gradient magnitude is performed along the gradient direction to obtain a refined edge response map; Adaptive double-threshold segmentation is applied to the refined edge response map to output a binary edge map; Morphological processing and connected component filtering are performed on the binary edge graph to form a cleaned edge graph; Extract a set of closed contours from the cleanup edge map, calculate the particle area of ​​each contour in the set, obtain the equivalent diameter based on the particle area, and then statistically analyze the particle size distribution and real-time particle size.

4. The method for controlling the CVD growth of silicon carbide particles based on multimodal sensing as described in claim 1, characterized in that, The hybrid reasoning mechanism that integrates deep reinforcement learning and fuzzy logic is as follows: Based on particle size deviation, particle size standard deviation, crystal form characterization data, process parameters, energy consumption, deposition rate, and structural state vectors of the production stage; The sensitivity weight vector for the compensation channel is obtained by inferring the state vector based on the dynamic sensitivity weight model. The multi-objective optimization weights are dynamically determined based on the production stage, and the particle size deviation, energy consumption and deposition rate are dynamically weighed in the multi-objective comprehensive cost function. Generate prior constraint or compensation channel preference vectors using fuzzy logic reasoning; Actions are selected under the compensation channel preference vector or prior constraints, and the optimal compensation strategy is generated through deep reinforcement learning.

5. The method for controlling the CVD growth of silicon carbide particles based on multimodal sensing as described in claim 4, characterized in that, The dynamic sensitivity weight model is constructed based on a gradient boosting decision tree, yielding a sensitivity weight vector for the compensation channel, expressed as: in, For the sensitivity weight vector, , , The relative sensitivity of temperature, pressure, and flow ratio to particle size deviation correction is characterized, respectively.

6. The method for restoring and enhancing low-quality character images of nuclear power plant fuel assemblies as described in claim 4, characterized in that, The action is defined as: in, For action, For temperature regulation, For pressure regulation, This is the flow rate adjustment amount.

7. The method for controlling the CVD growth of silicon carbide particles based on multimodal sensing as described in claim 1, characterized in that, The multi-objective integrated cost function is expressed as follows: in, For particle size deviation, For energy consumption, For the deposition rate, For particle size standard deviation, , , To follow the production stage Dynamically changing weights , The weights of the particle size dispersion and crystal form penalty terms are given. The crystal form threshold. As a crystal form index, This indicates a penalty for failing to meet crystal form standards.

8. A silicon carbide particle CVD growth control system based on multimodal sensing, characterized in that, include: The multi-source data acquisition module is configured to: synchronously acquire multimodal data inside the reaction chamber; the multimodal data includes particle morphology images, process parameters, and crystal form characterization data; The particle size recognition module is configured to: recognize the particle morphology image to obtain the real-time particle size; The compensation decision module is configured to: calculate the particle size deviation based on real-time particle size, calculate the crystal form optimization flag based on crystal form characterization data, and if the particle size deviation is greater than the preset particle size deviation threshold or the crystal form optimization flag is true, trigger the improved AI decision engine to make a compensation decision, obtain the optimal adjustment amount of the process parameters and execute it. The improved AI decision engine employs a hybrid reasoning mechanism that combines deep reinforcement learning and fuzzy logic to construct a multi-objective comprehensive cost function, with minimizing the multi-objective comprehensive cost function as the optimization objective.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the multimodal sensing-based silicon carbide particle CVD growth control method as described in any one of claims 1-7.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the multimodal sensing-based silicon carbide particle CVD growth control method as described in any one of claims 1-7.

Images