Quality verification method and system in silicon carbide purification process based on deep model

Abstract

The present invention discloses a quality verification method and system for silicon carbide purification processing based on a deep model, which relates to the field of silicon carbide inspection. The quality verification method for silicon carbide purification processing based on a deep model includes the following steps: classifying monitoring data to obtain classified data; obtaining association rules related to silicon carbide product defects; based on a deep learning model, combining the obtained association rules and parameter features in the classified data to predict the defect type and probability of the silicon carbide product; quantifying the predicted defect type and probability of the silicon carbide product to obtain a quantitative score, and combining the quantitative score with a risk index to calculate an overall risk score and construct a risk profile. The present invention combines the characteristics of the association rules and the characteristics of the classified data to improve the model's prediction accuracy for the defect type and probability of the silicon carbide product, quantifies the prediction results into a risk score, and combines them with other risk indicators to provide an intuitive quality verification evaluation indicator.

Description

Technical Field

[0001] The present invention relates to the field of silicon carbide inspection, and in particular to a quality verification method and system in silicon carbide purification processing based on a depth model. Background Art

[0002] Silicon carbide is a compound composed of silicon and carbon, bonded together by covalent bonds. It is widely used in various fields, including as an abrasive, as a high-temperature resistant material in the semiconductor industry, and as a high-temperature, high-pressure, and high-frequency semiconductor material in power electronics. Silicon carbide purification is the process of removing impurities (such as metallic impurities, oxides, and nitrides) from natural or artificial silicon carbide to obtain high-purity silicon carbide. This process is crucial to improving the application of silicon carbide in advanced electronic and optoelectronic devices. There are various purification methods, including physical vapor transport (PVT), chemical vapor deposition (CVD), and sodium salt purification.

[0003] Deep model-based quality verification in silicon carbide purification processes involves using deep learning models to evaluate and verify the quality control of the silicon carbide purification process. This involves identifying impurities, cracks, crystal structure defects, and other issues to assess the effectiveness of the purification process.

[0004] For example, Chinese patent 201910694877.8 discloses a method for evaluating the quality of high-purity silicon carbide powder. Based on the principle of spectrophotometry and the differences in the apparent color of the high-purity silicon carbide powder caused by various impurities in the powder, the method creatively constructs a quality factor calculation formula, which can semi-quantitatively analyze the nitrogen content in the silicon carbide powder. However, the above method still has the following shortcomings: spectrophotometry may not be easy to automate on a large scale, especially in a continuous production environment, and it cannot learn and identify complex patterns in the data and provide a more comprehensive and in-depth analysis.

[0005] Currently, no effective solutions have been proposed for the problems in related technologies. Summary of the Invention

[0006] In response to the problems in the related art, the present invention proposes a quality verification method and system for silicon carbide purification processing based on a depth model to overcome the above-mentioned technical problems existing in the existing related art.

[0007] To this end, the specific technical solutions adopted in the present invention are as follows:

[0008] According to one aspect of the present invention, a method for verifying the quality of silicon carbide during purification processing based on a depth model is provided. The method for verifying the quality of silicon carbide during purification processing based on a depth model comprises the following steps:

[0009] S1. Acquire monitoring data during silicon carbide purification processing, and classify the monitoring data according to different reaction periods to obtain classified data.

[0010] S2. Use association rule learning algorithm to analyze the classified data and obtain association rules related to silicon carbide product defects.

[0011] S3. Based on the deep learning model, combined with the acquired association rules and parameter features in the classification data, the defect type and probability of silicon carbide products are predicted.

[0012] S4. Quantify the defect types and probabilities of silicon carbide products and obtain a quantitative score. Combine the quantitative score with the risk index to calculate the overall risk score and construct a risk profile.

[0013] Among them, based on the deep learning model, combined with the acquired association rules and parameter features in the classification data, the defect type and probability of silicon carbide products are predicted, including the following steps:

[0014] S31. Convert association rules into numerical features for deep learning models;

[0015] S32, combining parameter features in the classification data with the association rule coding to form a comprehensive feature set;

[0016] S33. Collect labeled data and train a deep learning model using a comprehensive feature set and labeled data. At the same time, predict the defect type and probability of silicon carbide products based on the trained deep learning model.

[0017] Furthermore, obtaining monitoring data during the silicon carbide purification process and classifying the monitoring data according to different reaction periods, and obtaining the classified data includes the following steps:

[0018] S11. Obtain monitoring data, including temperature, pressure and impurity content parameters;

[0019] S12. Clean monitoring data, handle anomalies and missing values, and standardize data;

[0020] S13. Sub-classify the monitoring data according to different reaction stages.

[0021] Furthermore, using an association rule learning algorithm to analyze the classified data and obtain association rules related to silicon carbide product defects includes the following steps:

[0022] S21. Build and update the data dictionary based on hashing technology;

[0023] S22. Process the data dictionary based on transaction compression technology to generate a compressed transaction data set;

[0024] S23. Based on the number of occurrences of each classification data in the data dictionary, create a head pointer table, sort it in descending order of frequency, and remove parameters that do not meet the minimum support threshold from the head pointer table;

[0025] S24, constructing a frequent pattern tree, mining frequent item sets from the frequent pattern tree, and generating association rules based on the mined frequent item sets and according to the confidence level;

[0026] S25. Verify the association rules using new classification data.

[0027] Furthermore, building and updating the data dictionary based on hashing technology includes the following steps:

[0028] S211. Create an empty data dictionary. The structure of the data dictionary is [key:value], where key represents the parameter name and value represents the frequency of the parameter.

[0029] S212. Construct a hash function and map each parameter name to a hash value, where the hash value is used to determine the location where the parameter name is stored in the hash table.

[0030] S213, initialize the hash table, the hash table size is set according to the expected number of parameters and load factor;

[0031] S214, traversing the classified data one by one, and using a hash function to calculate a hash value of each parameter in the classified data;

[0032] S215. Find the corresponding storage location in the hash table according to the hash value. If the storage location is free, create a new entry in the storage location, use the parameter name as the key, and initialize the frequency to 1. If the storage location is occupied, update the frequency of the corresponding entry and increase the frequency by 1.

[0033] S216. If a hash conflict is encountered in the hash table, a conflict resolution strategy is adopted;

[0034] S217: After completing the traversal and hashing of all classified data, the hash table is optimized.

[0035] Furthermore, a frequent pattern tree is constructed, and frequent item sets are mined from the frequent pattern tree. The association rules are generated based on the mined frequent item sets and the confidence level, including the following steps:

[0036] S241. Create a root node of the frequent pattern tree and set it to null;

[0037] S242, traverse the compressed transaction data set, for each parameter in each transaction, if the parameter already exists in the frequent pattern tree, update the count of the corresponding node; if the parameter does not exist in the frequent pattern tree, create a new child node;

[0038] S243, obtaining parameter combinations that frequently appear in the frequent pattern tree to obtain frequent item sets;

[0039] S244. Use the mined frequent item sets and generate association rules based on the confidence level;

[0040] The calculation formula for support is:

[0041] ;

[0042] The confidence calculation formula is:

[0043] ;

[0044] A and B both represent item sets;

[0045] A∪B represents the union of item set A and item set B;

[0046] count(A∪B) represents the number of transactions in which item sets A and B appear simultaneously;

[0047] TotalCount indicates the total number of transactions;

[0048] count(A) represents the number of transactions in which item set A appears.

[0049] Furthermore, collecting labeled data and training a deep learning model using the comprehensive feature set and labeled data, and predicting the defect type and probability of silicon carbide products based on the trained deep learning model include the following steps:

[0050] S331, converting the comprehensive feature set and label data into a vector;

[0051] S332, fusing the convolutional neural network with the long short-term memory network to obtain a fusion network model;

[0052] S333, dividing the comprehensive feature set and label data converted into vectors into a training set and a test set;

[0053] S334, using the training set to train the fusion network model, and using the cross-validation method to optimize the performance of the fusion network model, while using the test set to evaluate the performance of the fusion network model;

[0054] S335. Use the trained fusion network model to predict defects on the new comprehensive feature set and obtain the probability of each defect type.

[0055] Furthermore, the defect types and probabilities of the predicted silicon carbide products are quantified to obtain a quantitative score, and the quantitative score is combined with the risk index to calculate the overall risk score, including the following steps:

[0056] S41. Pre-set a corresponding weight for each defect type, select the defect type with the highest probability, normalize the probability, and multiply the normalized probability by the corresponding weight to obtain a quantitative score;

[0057] S42. Weighting and quantifying risk indicators;

[0058] S43. Calculate the total risk score based on the quantitative scores and the quantified risk indicators.

[0059] Furthermore, the total risk score is calculated as follows:

[0060] ;

[0061] Where, RS represents the overall risk score, Q Indicates a quantitative score;

[0062] m Indicates the number of risk indicators;

[0063] W i Indicates the i The weight coefficient of each risk indicator, Z i Represents the quantized i risk indicators.

[0064] Furthermore, when constructing a risk profile, the quality verification results in the silicon carbide purification process are classified into different risk levels based on the total risk score, and the risk profile is drawn by color, and the total risk score and risk profile are displayed in real time.

[0065] According to another aspect of the present invention, a quality verification system for silicon carbide purification processing based on a deep model is provided, and the quality verification system for silicon carbide purification processing based on a deep model includes a classification data acquisition module, an association rule acquisition module, a deep learning module and a risk quantification module; wherein, the classification data acquisition module is connected to the association rule acquisition module, the association rule acquisition module is connected to the deep learning module, and the deep learning module is connected to the risk quantification module.

[0066] The classification data acquisition module is used to acquire monitoring data during the silicon carbide purification process and classify the monitoring data according to different reaction periods to obtain classified data.

[0067] The association rule acquisition module is used to analyze the classification data using the association rule learning algorithm to obtain association rules related to silicon carbide product defects.

[0068] The deep learning module is used to predict the defect type and probability of silicon carbide products based on the deep learning model, combined with the acquired association rules and parameter features in the classification data.

[0069] The risk quantification module is used to quantify the defect types and probabilities of predicted silicon carbide products, obtain quantitative scores, and combine the quantitative scores with risk indicators to calculate the overall risk score and construct a risk profile.

[0070] The beneficial effects of the present invention are:

[0071] (1) The deep model-based quality verification method and system for silicon carbide purification provided by the present invention manages and analyzes data in detail according to different reaction stages, breaking down the complex purification process into a series of simpler and more manageable stages, providing a foundation for subsequent data analysis and defect prediction. By using association rule learning algorithms, the hidden laws and patterns in the classified data are discovered, especially those related to silicon carbide product defects. This helps to understand the conditions and causes of defects, which is a key step in improving production processes and enhancing product quality.

[0072] (2) The present invention combines deep learning models with association rules and parameter features of classified data to more accurately predict the defect types and probabilities of silicon carbide products. Compared with traditional methods, association rules can usually mine potential patterns and correlations in data. Adding them to the original features can increase the dimension and richness of the features, helping the deep learning model to better grasp the inherent structure and complexity of the data. Combining the characteristics of association rules with the characteristics of classified data allows the model to capture more effective information, thereby improving the model's prediction accuracy for the defect types and probabilities of silicon carbide products. Association rules can help us understand how specific input features affect the output, thereby improving the interpretability of the model. The method of the present invention can be used in automated production processes.

[0073] (3) The present invention quantifies the prediction results into risk scores and combines them with other risk indicators to calculate the overall risk score and construct a risk profile, which can better understand and quantify the risks in the production process and provide an intuitive quality verification evaluation indicator. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0075] Figure 1is a flow chart of a quality verification method in a silicon carbide purification process based on a depth model according to an embodiment of the present invention;

[0076] Figure 2 4 is a principle block diagram of a quality verification system in a silicon carbide purification process based on a depth model according to an embodiment of the present invention.

[0077] In the picture:

[0078] 1. Classification data acquisition module; 2. Association rule acquisition module; 3. Deep learning module; 4. Risk quantification module. DETAILED DESCRIPTION

[0079] To further illustrate each embodiment, the present invention provides drawings, which are part of the disclosure of the present invention. They are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. By referring to these contents, ordinary technicians in this field should be able to understand other possible implementation methods and advantages of the present invention. The components in the figures are not drawn to scale, and similar component symbols are generally used to represent similar components.

[0080] According to an embodiment of the present invention, a method and system for quality verification in silicon carbide purification processing based on a depth model are provided.

[0081] The present invention will now be further described with reference to the accompanying drawings and specific embodiments. Figure 1 As shown, according to one embodiment of the present invention, a method for verifying the quality of silicon carbide during purification processing based on a depth model is provided. The method for verifying the quality of silicon carbide during purification processing based on a depth model comprises the following steps:

[0082] S1. Acquire monitoring data during silicon carbide purification processing, and classify the monitoring data according to different reaction periods to obtain classified data.

[0083] In a further embodiment, obtaining monitoring data during the silicon carbide purification process and classifying the monitoring data according to different reaction periods to obtain the classified data includes the following steps:

[0084] S11. Obtain monitoring data, including temperature, pressure, and impurity content parameters; these parameters are crucial for controlling the silicon carbide purification process. Temperature and pressure monitoring ensures the ideal reaction environment, while impurity content monitoring helps evaluate purification results and adjust process parameters in a timely manner.

[0085] S12. Clean monitoring data, handle anomalies and missing values, and standardize data to ensure consistency and comparability during model training.

[0086] S13. Sub-classifying monitoring data by reaction stage (e.g., heating, holding, cooling, etc.) can help the model better understand how parameter changes during different stages affect product quality. For example, temperature and pressure changes during the heating stage significantly impact impurity removal efficiency, while stability during the holding stage influences the crystallization quality of the final product. Sub-classifying data for each stage can help build more refined models, leading to a deeper understanding and control of the entire purification process.

[0087] S2. Use association rule learning algorithms to analyze the categorized data and obtain association rules related to silicon carbide product defects. Association rules are used to discover the relationships between various factors that lead to product defects.

[0088] In a further embodiment, analyzing the classified data using an association rule learning algorithm to obtain association rules related to silicon carbide product defects includes the following steps:

[0089] S21. Build and update a data dictionary based on hashing techniques. Hashing techniques are used to quickly find and retrieve data. Building a data dictionary involves hashing the items in the original dataset and storing them as key-value pairs for efficient lookup and update.

[0090] S22. Process the data dictionary based on transaction compression technology to generate a compressed transaction data set; transaction compression technology is used to reduce the size of the data set and simplify subsequent processing steps by merging or omitting certain data.

[0091] S23. Create a header table based on the number of occurrences of each categorical data item in the data dictionary and sort it in descending order by frequency. Items that do not meet the minimum support threshold are removed from the header table. The header table stores each item and its frequency in the dataset, with items sorted in descending order by frequency. The minimum support is a user-defined threshold that indicates the minimum number of times an item set must appear to be considered frequent. Items that do not meet this threshold are removed, which reduces the size of the frequent pattern tree and improves the efficiency of mining frequent itemsets.

[0092] S24. Construct a frequent pattern tree (FP tree) and mine frequent item sets from the frequent pattern tree. Generate association rules based on the mined frequent item sets and according to the confidence level. FP tree is a compressed data structure used to store frequent item sets. Once the FP tree is constructed, specific algorithms (such as FP-Growth algorithm) can be used to quickly mine frequent item sets.

[0093] S25. Verify the association rules using new classified data. If the association rules perform well on the new data, they can be used for further analysis or decision support.

[0094] In a further embodiment, building and updating a data dictionary based on hashing technology includes the following steps:

[0095] S211. Create an empty data dictionary. The structure of the data dictionary is [key:value], where key represents the parameter name and value represents the frequency of the parameter.

[0096] S212. Construct a hash function and map each parameter name to a hash value, and the hash value is used to determine the storage location of the parameter name in the hash table; wherein the hash function is a function that maps an input (here, the parameter name) to a numerical value (hash value).

[0097] S213. Initialize a hash table (or hash table). The size of the hash table is set according to the expected number of parameters and the load factor (ie, the ratio of the number of elements to the number of positions in the table).

[0098] S214 , traversing the classified data one by one, and using a hash function to calculate a hash value of each parameter in the classified data, to ensure that each parameter is correctly mapped into the hash table.

[0099] S215. Find the corresponding storage location in the hash table according to the hash value. If the storage location is free, create a new entry in the storage location, use the parameter name as the key, and initialize the frequency to 1. If the storage location is occupied, update the frequency of the corresponding entry and increase the frequency by 1.

[0100] S216. If a hash conflict is encountered in the hash table (i.e., different parameter names are mapped to the same hash value), a conflict resolution strategy is adopted, such as the chain address method or the open addressing method.

[0101] S217. After completing the traversal and hashing of all classified data, the hash table is optimized, such as adjusting the size or rehashing to reduce conflicts.

[0102] In a further embodiment, constructing a frequent pattern tree, mining frequent item sets from the frequent pattern tree, and generating association rules based on the mined frequent item sets and confidence levels include the following steps:

[0103] S241. Create a root node of the frequent pattern tree and set it to null. The root node will serve as the starting point of all subsequent frequent item set paths.

[0104] S242. Traverse the compressed transaction data set. For each parameter in each transaction, if the parameter already exists in the frequent pattern tree, update the count of the corresponding node. If the parameter does not exist in the frequent pattern tree, create a new child node. Repeat this for all transactions in the data set, and finally form a frequent pattern tree containing all frequent item set information.

[0105] S243. Obtain parameter combinations that frequently appear in the frequent pattern tree to obtain frequent item sets.

[0106] S244. Use the mined frequent item sets and generate association rules based on confidence. The association rules will indicate which parameters often appear together and may be related to silicon carbide product defects.

[0107] The calculation formula for support is:

[0108] ;

[0109] The confidence calculation formula is:

[0110] ;

[0111] A and B both represent item sets, which are the set of one set of items and the set of another set of items in the database. A and B are disjoint;

[0112] A∪B represents the union of item set A and item set B, which means the set of transactions that contains all items in both A and B;

[0113] count(A∪B) represents the number of transactions in which item sets A and B appear simultaneously;

[0114] TotalCount indicates the total number of transactions;

[0115] count(A) represents the number of transactions in which item set A appears.

[0116] In order to facilitate understanding of the above technical solutions of the present invention, the working principle or operation mode of the present invention in actual process is described in detail below.

[0117] Acquire monitoring data, including temperature, pressure, and impurity content parameters. This data is collected during the silicon carbide purification process. The monitoring data is shown in Table 1.

[0118] Table 1 Monitoring data

[0119]

[0120] If the temperature exceeds 1600°C or is below 1400°C, treat it as an outlier and exclude it. If data for a timestamp is missing, use the average of the previous and next time points to fill in the gaps. Convert all parameters to z-scores. Subclassify the data according to the reaction phase (preheating, reaction, cooling). Use hashing technology to create a data dictionary to store parameter names and frequency of occurrence. Perform transaction compression on the data dictionary to reduce the size of the dataset. Create a header pointer table and remove parameters that do not meet the minimum support threshold. Construct a frequent pattern tree and mine frequent itemsets from it. Use new classified data to verify the accuracy of the association rules.

[0121] Frequent itemsets: {temperature: 1500, pressure: 10}, {temperature: 1550, pressure: 12}.

[0122] Association rule: If temperature = 1500 and pressure = 10, then the probability of impurity content increasing is 0.7.

[0123] S3. Based on the deep learning model, combined with the acquired association rules and parameter features in the classification data, the defect type and probability of silicon carbide products are predicted.

[0124] Among them, based on the deep learning model, combined with the acquired association rules and parameter features in the classification data, the defect type and probability of silicon carbide products are predicted, including the following steps:

[0125] S31. Convert association rules into numerical features (such as sequence coding, hash coding, etc.) for deep learning models.

[0126] S32. Combine the parameter features in the categorical data with the association rule encoding to form a comprehensive feature set. This feature set, which includes both the features of the original data and the new features representing the association rules, is provided as input to the deep learning model. This provides the model with richer information and helps it understand complex patterns in the data.

[0127] S33. Collect labeled data and train a deep learning model using the combined feature set and labeled data. The trained deep learning model then predicts the defect type and probability of silicon carbide products. During training, the model learns how to predict the probability of a defect type based on the input feature set. Once trained, the model can be used to predict defect types for new product samples.

[0128] In a further embodiment, collecting labeled data, training a deep learning model using the integrated feature set and labeled data, and predicting defect types and probabilities of silicon carbide products based on the trained deep learning model include the following steps:

[0129] S331. Convert the comprehensive feature set and label data into vectors, and use word embedding and other methods to convert them into vector form.

[0130] S332. Fuse a convolutional neural network with a long short-term memory network to create a fused network model. Use the CNN to extract local features of the data. Characters are converted into fixed-length vectors through an embedding layer, and then the convolutional layers are used to extract local patterns and features from these word vectors. Use LSTM to process the sequential nature of the data and extract global features. Adaptively adjust the architecture of the LSTM-CNN model based on the characteristics of silicon carbide product defects, such as by changing the number of convolutional layers, the number of LSTM units, or the configuration of fully connected layers.

[0131] S333. Divide the comprehensive feature set and label data converted into vectors into a training set and a test set.

[0132] S334. Use the training set to train the fusion network model, and use the cross-validation method to optimize the performance of the fusion network model. At the same time, use the test set to evaluate the performance of the fusion network model, paying attention to indicators such as accuracy, recall rate and F1 score.

[0133] S335. Use the trained fusion network model (i.e., deep learning model) to predict defects on the new comprehensive feature set and obtain the probability of each defect type.

[0134] S4. Quantify the predicted defect types and probabilities of silicon carbide products to obtain a quantitative score. Combine the quantitative score with risk indicators (such as temperature deviation, pressure deviation, and impurity content) to calculate the overall risk score and construct a risk profile.

[0135] In a further embodiment, quantifying the predicted defect types and probabilities of the silicon carbide product to obtain a quantitative score, and combining the quantitative score with the risk index to calculate an overall risk score includes the following steps:

[0136] S41. Pre-set a corresponding weight for each defect type, select the defect type with the highest probability, and normalize the probability so that its value is within a specific range, such as 0 to 1. Multiply the normalized probability by the corresponding weight to obtain a quantitative score. The quantitative score reflects the risk of the silicon carbide product having the defect.

[0137] S42. Weighting and quantifying risk indicators;

[0138] Among them, the risk indicators are weighted to reflect their impact on the risk score. The risk indicators are quantified into numerical values ​​for easy calculation. i =f(R i ), where R i It isi The actual measured value of the risk indicator, Z i is the quantized value, and f is the quantization function, which is designed according to the actual situation, for example, using linear mapping or normalization method.

[0139] S43. Calculate the total risk score based on the quantitative scores and the quantified risk indicators.

[0140] In a further embodiment, the total risk score is calculated as:

[0141] ;

[0142] Where, RS represents the overall risk score, Q Indicates a quantitative score;

[0143] m Indicates the number of risk indicators;

[0144] W i Indicates the i The weight coefficient of each risk indicator, Z i Represents the quantized i risk indicators.

[0145] In a further embodiment, when constructing a risk portrait, the quality verification results in the silicon carbide purification process are classified into different risk levels based on the total risk score, and the risk portrait is drawn by color so that people can intuitively see the distribution and size of the risk, and the total risk score and risk portrait are displayed in real time.

[0146] In order to facilitate understanding of the above technical solutions of the present invention, the working principle or operation mode of the present invention in actual process is described in detail below.

[0147] Set the weight of microcracks to 0.7 and the weight of particle impurities to 0.8. Select particle impurities as the defect type with the highest probability. The probability of particle impurities is 0.8, which is normalized to 0.8. For example, the quantized score of particle impurities is 0.64 (0.8 * 0.8).

[0148] The weight of temperature deviation is 0.5, the weight of pressure deviation is 0.3, and the weight of impurity content is 0.2.

[0149] Quantify the risk indicators: temperature deviation is 0.1, pressure deviation is 0.05, and impurity content is 0.1. Define a linear mapping function f(x) = x, and then calculate the quantitative values ​​of temperature deviation equal to 0.1, pressure deviation equal to 0.05, and impurity content equal to 0.1. Add all weighted quantitative scores to obtain the total risk indicator quantitative score of 0.085. When drawing a risk profile, low risk is green, medium risk is yellow, and high risk is red.

[0150] like Figure 2 As shown, according to another embodiment of the present invention, a quality verification system for silicon carbide purification processing based on a deep model is provided, and the quality verification system for silicon carbide purification processing based on a deep model includes a classification data acquisition module 1, an association rule acquisition module 2, a deep learning module 3 and a risk quantification module 4; wherein, the classification data acquisition module 1 is connected to the association rule acquisition module 2, the association rule acquisition module 2 is connected to the deep learning module 3, and the deep learning module 3 is connected to the risk quantification module 4.

[0151] The classified data acquisition module 1 is used to acquire monitoring data during the silicon carbide purification process and classify the monitoring data according to different reaction periods to obtain classified data.

[0152] The association rule acquisition module 2 is used to analyze the classified data using an association rule learning algorithm to obtain association rules related to silicon carbide product defects.

[0153] Deep learning module 3 is used to predict the defect type and probability of silicon carbide products based on the deep learning model, combined with the acquired association rules and parameter features in the classification data.

[0154] The risk quantification module 4 is used to quantify the defect types and probabilities of the predicted silicon carbide products, obtain a quantitative score, and combine the quantitative score with the risk index to calculate the overall risk score and construct a risk profile.

[0155] In summary, the quality verification method and system for silicon carbide purification processing based on a deep model provided by the present invention manages and analyzes data in detail according to different reaction stages, decomposes the complex purification process into a series of simpler and more manageable stages, and provides a basis for subsequent data analysis and defect prediction. By using an association rule learning algorithm, the hidden laws and patterns in the classified data are discovered, especially the rules related to silicon carbide product defects, which helps to understand the conditions and causes of defects, which is a key step in improving the production process and improving product quality. The present invention combines a deep learning model with association rules and parameter features of classified data to more accurately predict the defect type and probability of silicon carbide products. Compared with traditional methods, association rules can usually mine the potential laws and correlations in the data. Adding them to the original features can increase the dimension and richness of the features, which helps the deep learning model better grasp the inherent structure and complexity of the data. Combining the characteristics of association rules and the characteristics of classified data enables the model to capture more effective information, thereby improving the model's prediction accuracy for the defect type and probability of silicon carbide products. Association rules can help us understand how specific input features affect the output, thereby improving the interpretability of the model. The method of the present invention can be applied to automated production processes. The method quantifies the prediction results into a risk score and combines it with other risk indicators to calculate an overall risk score and construct a risk profile. This allows for a better understanding and quantification of risks in the production process, providing an intuitive quality verification and assessment metric.

[0156] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A quality verification method for silicon carbide purification based on a depth model, characterized in that: The quality verification method for silicon carbide purification process based on the depth model includes the following steps: S1. Acquire monitoring data during silicon carbide purification processing, and classify the monitoring data according to different reaction periods to obtain classified data; S2. Analyze the classified data using an association rule learning algorithm to obtain association rules related to silicon carbide product defects; S3. Convert association rules into numerical features for deep learning models; combine parameter features in the classification data with the association rule encoding to form a comprehensive feature set; collect labeled data and train the deep learning model using the comprehensive feature set and labeled data; and predict the defect type and probability of silicon carbide products based on the trained deep learning model, specifically including: Convert the comprehensive feature set and label data into vectors; fuse the convolutional neural network with the long short-term memory network to obtain a fusion network model; divide the comprehensive feature set and label data converted into vectors into a training set and a test set; use the training set to train the fusion network model, and use cross-validation to optimize the performance of the fusion network model, while using the test set to evaluate the performance of the fusion network model; use the trained fusion network model to predict defects on the new comprehensive feature set and obtain the probability of each defect type; S4. Quantify the predicted defect types and probabilities of silicon carbide products to obtain a quantitative score. Combine the quantitative score with the risk index to calculate the overall risk score and construct a risk profile. Specifically, this includes: S41. Set a corresponding weight for each defect type in advance, select the defect type with the highest probability, standardize the probability, and multiply the standardized probability by the corresponding weight to obtain a quantitative score; S42. Weight and quantify the risk indicators; S43. Calculate the total risk score based on the quantitative score and the quantified risk indicators.

2. The method for quality verification in silicon carbide purification process based on a depth model according to claim 1, characterized in that: The method of obtaining monitoring data during the silicon carbide purification process and classifying the monitoring data according to different reaction periods to obtain classified data includes the following steps: S11. Obtain monitoring data, including temperature, pressure and impurity content parameters; S12. Clean monitoring data, handle anomalies and missing values, and standardize data; S13. Sub-classify the monitoring data according to different reaction stages.

3. The method for quality verification in silicon carbide purification process based on a depth model according to claim 2, characterized in that: The method of analyzing the classified data using the association rule learning algorithm to obtain association rules related to silicon carbide product defects includes the following steps: S21. Build and update the data dictionary based on hashing technology; S22. Process the data dictionary based on transaction compression technology to generate a compressed transaction data set; S23. Based on the number of occurrences of each classification data in the data dictionary, create a head pointer table, sort it in descending order of frequency, and remove parameters that do not meet the minimum support threshold from the head pointer table; S24, constructing a frequent pattern tree, mining frequent item sets from the frequent pattern tree, and generating association rules based on the mined frequent item sets and according to the confidence level; S25. Verify the association rules using new classification data.

4. The method for quality verification in silicon carbide purification process based on a depth model according to claim 3, characterized in that: The method of constructing and updating the data dictionary based on the hashing technology includes the following steps: S211. Create an empty data dictionary. The structure of the data dictionary is [key:value], where key represents the parameter name and value represents the frequency of the parameter. S212. Construct a hash function and map each parameter name to a hash value, where the hash value is used to determine the location where the parameter name is stored in the hash table. S213, initialize the hash table, the hash table size is set according to the expected number of parameters and load factor; S214, traversing the classified data one by one, and using a hash function to calculate a hash value of each parameter in the classified data; S215. Find the corresponding storage location in the hash table according to the hash value. If the storage location is free, create a new entry in the storage location, use the parameter name as the key, and initialize the frequency to 1. If the storage location is occupied, update the frequency of the corresponding entry and increase the frequency by 1. S216. If a hash conflict is encountered in the hash table, a conflict resolution strategy is adopted; S217: After completing the traversal and hashing of all classified data, the hash table is optimized.

5. The method for quality verification in silicon carbide purification process based on a depth model according to claim 4, characterized in that: The process of constructing a frequent pattern tree, mining frequent item sets from the frequent pattern tree, and generating association rules based on the mined frequent item sets and confidence levels comprises the following steps: S241. Create a root node of the frequent pattern tree and set it to null; S242, traverse the compressed transaction data set, for each parameter in each transaction, if the parameter already exists in the frequent pattern tree, update the count of the corresponding node; if the parameter does not exist in the frequent pattern tree, create a new child node; S243, obtaining parameter combinations that frequently appear in the frequent pattern tree to obtain frequent item sets; S244. Use the mined frequent item sets and generate association rules based on the confidence level; The calculation formula for support is: ； The confidence calculation formula is: ； A and B both represent item sets; A∪B represents the union of item set A and item set B; count(A∪B) represents the number of transactions in which item sets A and B appear simultaneously; TotalCount indicates the total number of transactions; count(A) represents the number of transactions in which item set A appears.

6. The method for quality verification in silicon carbide purification process based on a depth model according to claim 1, characterized in that: The overall risk score is calculated as: ； Where, RS represents the overall risk score, Q Indicates a quantitative score; m Indicates the number of risk indicators; W i Indicates the i The weight coefficient of each risk indicator, Z i Represents the quantized i risk indicators.

7. The method for quality verification in silicon carbide purification process based on a depth model according to claim 1, characterized in that: When constructing the risk portrait, the quality verification results in the silicon carbide purification process are classified into different risk levels based on the total risk score, and the risk portrait is drawn according to color, and the total risk score and risk portrait are displayed in real time.

8. A quality verification system for silicon carbide purification process based on a depth model, used to implement the quality verification method for silicon carbide purification process based on a depth model according to any one of claims 1 to 7, characterized in that: The quality verification system for silicon carbide purification processing based on deep models includes a classification data acquisition module, an association rule acquisition module, a deep learning module and a risk quantification module; Wherein, the classification data acquisition module is connected to the association rule acquisition module, the association rule acquisition module is connected to the deep learning module, and the deep learning module is connected to the risk quantification module; The classification data acquisition module is used to acquire monitoring data during the silicon carbide purification process and classify the monitoring data according to different reaction periods to obtain classification data; The association rule acquisition module is used to analyze the classification data using an association rule learning algorithm to obtain association rules related to silicon carbide product defects; The deep learning module is used to predict the defect type and probability of silicon carbide products based on the deep learning model in combination with the acquired association rules and parameter features in the classification data; The risk quantification module is used to quantify the defect types and probabilities of the predicted silicon carbide products, obtain a quantitative score, and combine the quantitative score with the risk index to calculate the overall risk score and construct a risk profile.

Images