A method and system for on-line monitoring and early warning of czochralski single crystal

Abstract

The application discloses a Czochralski single crystal real-time online monitoring and early warning method and system, belongs to the field of Czochralski single crystal production and manufacturing, and aims to solve the problems of inaccurate judgment, no closed loop in disposal, poor real-time performance and the like of traditional monitoring. The method collects three types of original data of crystal growth, broken line and quality, standardizes the original data, generates structured data through furnace table level sliding window statistics, judges the abnormal type and grade after extracting features in combination with a dynamic threshold library and a multi-parameter algorithm, pushes early warning information containing process adjustment suggestions in a graded and directional manner, strengthens monitoring after process adjustment, verifies the effect, forms a closed loop and updates the threshold library. The system contains six function modules, and relies on edge computing and other technologies to ensure the real-time performance of the whole process. The application improves the accuracy of abnormal judgment and disposal efficiency, realizes closed loop control of the whole production process, adapts to large-scale production of multiple furnaces, and guarantees the quality of crystal rods and production efficiency.

Description

Technical Field

[0001] This invention relates to the field of Czochralski single crystal production and manufacturing, specifically to an online monitoring and early warning method and system for Czochralski single crystals. Background Technology

[0002] In the field of Czochralski single crystal production, as a core process link in industries such as semiconductors and photovoltaics, the production process covers multiple key process stages, including crystal pulling, shoulder formation, equal diameter, and finishing. It involves many related parameters such as crystal rod diameter, furnace temperature, and resistivity. Moreover, with the large-scale development of the industry, the parallel operation of multiple furnaces has become the mainstream production mode, which places extremely high demands on the real-time monitoring of the production process and the rapid identification and handling of anomalies. However, the traditional Czochralski single crystal production monitoring model has many technical pain points. First, data collection lacks standardization and comprehensiveness, often involving scattered, single-dimensional data collection. Data from different furnaces lacks unified labeling rules, easily leading to missing data dimensions and inconsistent labels, creating potential for missed or incorrect anomaly detection and failing to meet the unified data management needs of large-scale multi-furnace operations. Second, data processing and anomaly detection lack scientific rigor. Raw data lacks systematic statistical analysis, and anomaly detection relies on manual experience, which is highly subjective and often uses single-parameter judgment logic, making it difficult to identify production anomalies caused by multiple factors, and lacks a unified anomaly level classification standard. Third, early warning notifications and responsibility matching lack specificity. There is no tiered early warning notification method, resulting in chaotic anomaly information dissemination, easily causing information interference from irrelevant personnel and hindering the response of key responsible persons. The problems are as follows: First, there is a lag in the process, and the lack of precise matching solutions for process adjustments requires responsible personnel to spend time analyzing and troubleshooting the direction of adjustments, which significantly reduces the efficiency of handling anomalies. Second, there is a lack of a complete closed-loop management system. After process adjustments, there are no objective and quantifiable standards for verifying the effects. The effects of adjustments are entirely based on manual judgment, and the judgment standards are not optimized and updated after anomaly handling, which easily leads to the recurrence of similar anomalies. At the same time, the traditional monitoring communication and data processing technology support is insufficient, the whole process has high latency, and the multi-threaded parallel processing capability is lacking, which cannot meet the technical requirements of real-time monitoring of multiple furnaces. The above problems ultimately lead to unstable crystal rod quality, frequent production accidents such as wire breakage, low production efficiency, and significant production losses for enterprises during the Czochralski single crystal production process. Therefore, there is an urgent need for an online monitoring and early warning system that can achieve real-time, standardized, and intelligent operation of the entire process to solve many of the drawbacks of the traditional monitoring mode. Summary of the Invention

[0003] The purpose of this invention is to provide an online monitoring and early warning method and system for Czochralski single crystals, thereby solving the problems mentioned in the background art.

[0004] This invention is achieved through the following technical solution: A real-time online monitoring and early warning method for Czochralski single crystals includes the following steps: S1. Collect raw data on crystal formation status, wire breakage, and quality inspection of each furnace in real time, and standardize and name them, then output the raw data with standardized names. S2. Furnace-level data statistics and standardization processing: After the raw data is classified by furnace number according to the standardized naming, it is statistically analyzed by sliding window to generate structured statistical data. S3. Extract features from structured statistical data to generate standardized feature data, retrieve the dynamic threshold library and use a multi-parameter combination judgment algorithm to complete the anomaly type identification and level classification, and output the standardized anomaly judgment result. S4. Based on the standardized anomaly judgment results, trigger the early warning push method according to the anomaly level, match the push responsibility according to the furnace and anomaly type, and push early warning information including suggested adjustment directions; S5. Enhanced monitoring after process adjustment: After the responsible person confirms receipt of the pushed early warning information, the enhanced monitoring of the corresponding furnace is started, the collection frequency is increased to continuously collect the characteristic data after process adjustment, and the real-time characteristic data after process adjustment is output. S6. Verification of Adjustment Effect and Closed-Loop Judgment: The effectiveness of the process adjustment is verified by comparing the feature data after the process adjustment with the judgment threshold range of the normal range of feature data in the dynamic threshold library. If the feature data is within the normal feature data range, the closed-loop judgment result is output; otherwise, the closed-loop judgment result of the adjustment is invalid is output.

[0005] Furthermore, In S1, the raw data on crystal formation status includes the crystal rod diameter, growth rate, furnace temperature, and argon flow rate at each process stage of the Czochralski single crystal furnace, including crystal introduction, shoulder formation, equal diameter formation, and finishing. The raw data related to wire breakage includes the historical number of wire breaks in the Czochralski single crystal furnace, the process stage in which the wire breakage occurred, the fluctuation of the crystal rod diameter within a preset time before the wire breakage, and the fluctuation data of the furnace temperature. The raw data for quality inspection includes the resistivity, oxygen content, and minority carrier lifetime data of the crystal rod.

[0006] Furthermore, In S2, the statistical analysis of the sliding window includes calculating the average value of crystallization state data, the frequency of wire breakage related data, and the change rate of quality inspection data within the window. The structured statistical data includes the crystallization state trend curve of the corresponding furnace, the wire breakage risk frequency report, and the quality index change report.

[0007] Furthermore, In S3, the feature extraction of structured statistical data includes: calculating the breakage rate of the corresponding furnace as a crystallization state feature; calculating the number of breakage warning triggers and the duration of diameter fluctuation standard deviation exceeding the standard in the equal diameter stage within a preset time period as breakage risk features; and calculating the average rate of change of resistivity, oxygen content, and minority carrier lifetime of the corresponding furnace within multiple consecutive sliding windows as quality index features. The standardized feature data is in the format of furnace number-feature type-feature value.

[0008] Furthermore, In S3, the dynamic threshold library pre-stores basic judgment thresholds for Czochralski single crystal production, including the judgment thresholds for the standard deviation of diameter fluctuation in the equal diameter stage, the rate of change of furnace temperature, the number of times the wire breakage warning is triggered, the duration of diameter fluctuation exceeding the standard, and the average value of the rate of change of quality indicators. The multi-parameter combination judgment algorithm uniformly divides anomalies into two levels: general anomalies and severe anomalies. Based on the comparison results of the characteristic values ​​of crystal formation state, wire breakage risk, and quality deterioration with the corresponding judgment thresholds, it completes the accurate judgment of the anomaly type and level.

[0009] Furthermore, In S4, the early warning push method corresponds one-to-one with the anomaly level: general anomalies trigger a pop-up window on the workstation computer + real-time message push to the enterprise, and the pop-up window and message continue until the responsible person confirms receipt; serious anomalies, in addition to the general anomaly push method, add an audible and visual alarm next to the furnace + SMS push to the responsible person's mobile phone, and the audible and visual alarm continues until the anomaly is cleared.

[0010] Furthermore, In S4, the person responsible for pushing the notification is matched with the level of the anomaly and the furnace platform: general anomalies are matched with the on-site operator and the process engineer of the corresponding furnace platform; severe anomalies are matched with the on-site operator, the process engineer of the corresponding furnace platform, the workshop technical supervisor, and the production administrator.

[0011] Furthermore, In S4, the suggested adjustment direction is generated based on the pre-stored mapping relationship between abnormality type, abnormality feature, and process adjustment scheme. According to the specific type and level of crystallization abnormality, wire breakage risk, and quality deterioration, the corresponding crystal rod growth and furnace environment adjustment strategies are matched.

[0012] Furthermore, In S6, the specific rules for closed-loop determination are as follows: if all real-time feature data after process adjustment within a preset time period returns to the judgment threshold range of the dynamic threshold library, the closed loop is determined to be successful, the corresponding furnace platform is marked as resolved and the normal data acquisition frequency is restored; if the feature data does not all return to the threshold range within a preset time period, the adjustment is determined to be invalid, the abnormality level is upgraded and a second early warning push is triggered, and continuous enhanced monitoring is carried out until the abnormality is resolved.

[0013] Furthermore, A real-time online monitoring and early warning system for Czochralski single crystals, used to implement the real-time online monitoring and early warning method for Czochralski single crystals as described above, characterized in that it includes: a multi-dimensional data acquisition module, a furnace-level automatic statistical analysis module, an anomaly intelligent judgment module, a targeted real-time early warning push module, a closed-loop handling and tracking module, and a real-time performance assurance module; The multi-dimensional data acquisition module establishes a gigabit network high-speed communication connection with the industrial control system of each Czochralski single crystal furnace to collect multi-dimensional raw data and complete standardized naming, and output standardized named raw data. The furnace-level automatic statistical analysis module is connected to the multi-dimensional data acquisition module to receive standardized named raw data, classify it according to the furnace number, and generate structured statistical data through sliding window statistical analysis. The anomaly intelligent judgment module is communicatively connected to the furnace-level automatic statistical analysis module. It has a built-in dynamic threshold library for extracting features from structured statistical data to generate standardized feature data. It also completes anomaly judgment through a multi-parameter combination judgment algorithm and outputs standardized anomaly judgment results. The targeted real-time early warning push module is communicatively connected to the anomaly intelligent judgment module. It is used to complete the hierarchical early warning triggering and push responsibility matching based on the standardized anomaly judgment results, and push early warning information to the corresponding responsible person's terminal. The closed-loop handling tracking module is connected to the directional real-time early warning push module and the anomaly intelligent judgment module. It has a built-in knowledge base and is used to start the furnace platform post-enhanced monitoring, collect real-time feature data after process adjustment, verify the adjustment effect and output the closed-loop judgment result. When the closed loop is successful, it stores the abnormal cases and updates and optimizes the dynamic threshold library, and feeds the optimized dynamic threshold library back to the anomaly intelligent judgment module. The real-time performance assurance module establishes communication connections with all other modules to provide real-time support for the entire system through edge computing, gigabit network high-speed communication, and multi-threaded parallel processing technology, ensuring that the total latency of the entire process from data acquisition to early warning push meets the requirements of real-time monitoring of Czochralski single crystal. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the process logic of the present invention; Figure 2 This is a schematic diagram of the system results. Detailed Implementation

[0015] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0016] See the example. Figures 1 to 2 : A real-time online monitoring and early warning method for Czochralski single crystals includes the following steps: S1. Collect raw data on crystal formation status, wire breakage, and quality inspection of each furnace in real time, and standardize and name them, then output the raw data with standardized names. S2. Furnace-level data statistics and standardization processing: After the raw data is classified by furnace number according to the standardized naming, it is statistically analyzed by sliding window to generate structured statistical data. S3. Extract features from structured statistical data to generate standardized feature data, retrieve the dynamic threshold library and use a multi-parameter combination judgment algorithm to complete the anomaly type identification and level classification, and output the standardized anomaly judgment result. S4. Based on the standardized anomaly judgment results, trigger the early warning push method according to the anomaly level, match the push responsibility according to the furnace and anomaly type, and push early warning information including suggested adjustment directions; S5. Enhanced monitoring after process adjustment: After the responsible person confirms receipt of the pushed early warning information, the enhanced monitoring of the corresponding furnace is started, the collection frequency is increased to continuously collect the characteristic data after process adjustment, and the real-time characteristic data after process adjustment is output. S6. Verification of Adjustment Effect and Closed-Loop Judgment: The effectiveness of the process adjustment is verified by comparing the feature data after the process adjustment with the judgment threshold range of the normal range of feature data in the dynamic threshold library. If the feature data is within the normal feature data range, the closed-loop judgment result is output; otherwise, the closed-loop judgment result of the adjustment is invalid is output.

[0017] Furthermore, In S1, the raw data on crystal formation status includes the crystal rod diameter, growth rate, furnace temperature, and argon flow rate at each process stage of the Czochralski single crystal furnace, including crystal introduction, shoulder formation, equal diameter formation, and finishing. The raw data related to wire breakage includes the historical number of wire breaks in the Czochralski single crystal furnace, the process stage in which the wire breakage occurred, the fluctuation of the crystal rod diameter within a preset time before the wire breakage, and the fluctuation data of the furnace temperature. The raw data for quality inspection includes the resistivity, oxygen content, and minority carrier lifetime data of the crystal rod.

[0018] The data collection system has formed a standardized and comprehensive collection system, which avoids the problems of missing data dimensions and inconsistent collection targets from the source. It provides comprehensive and standardized raw data support for subsequent furnace-level data statistical analysis, fundamentally reducing the probability of missed or misjudged anomalies due to incomplete data. At the same time, the unified collection targets also meet the unified data management needs of large-scale production of multiple furnaces. In practical implementation, this module establishes a gigabit network high-speed communication connection with the industrial control system of 126 Czochralski single crystal furnaces. It accurately collects core parameters such as crystal rod diameter, growth rate, furnace temperature, and argon flow rate at each process stage, including crystal pulling, shoulder formation, equal diameter, and finishing. It collects data related to wire breakage, including the number of wire breaks in the furnace, the process stage at which the wire breakage occurred, and the fluctuations in crystal rod diameter and furnace temperature in the 10 minutes before the wire breakage. Quality inspection data is collected every 10 minutes, including data on the resistivity, oxygen content, and minority carrier lifetime of the crystal rod. All collected raw data are standardized and named according to "furnace number-collection time-data type", which not only ensures the comprehensiveness of data collection for a single furnace, but also provides a unified identification rule for data from multiple furnaces, realizing efficient collection and orderly management of data from multiple furnaces under large-scale production.

[0019] Furthermore, In S2, the statistical analysis of the sliding window includes calculating the average value of crystallization state data, the frequency of wire breakage related data, and the change rate of quality inspection data within the window. The structured statistical data includes the crystallization state trend curve of the corresponding furnace, the wire breakage risk frequency report, and the quality index change report.

[0020] The standardized furnace-level data processing logic calculates the average value of crystallization status data, the frequency of breakage-related data, and the change rate of quality inspection data through a sliding window. This transforms scattered and irregular raw data into trend-based and regular statistical data. The structured data forms, such as crystallization status trend curves, breakage risk frequency reports, and quality indicator change reports, can intuitively and clearly present the changes in the production status of each furnace. This provides valuable and easily analyzable structured input for the subsequent feature extraction process, which aims to eliminate invalid information and extract core features. It avoids the disorganization of directly analyzing raw data and significantly improves the efficiency and relevance of data processing. In practical implementation, the system uses a fixed 5-minute sliding window to automatically perform statistical analysis on the standardized named raw data categorized by furnace number. It calculates the mean of crystallization status data, the frequency of occurrence of wire breakage-related data, and the rate of change of quality inspection data within each window. This generates exclusive crystallization status trend curves, wire breakage risk frequency reports, and quality indicator change reports for each of the 126 furnaces. These structured statistical data clearly show the changes in crystallization trends, fluctuations in wire breakage risk frequency, and dynamics of quality indicators for each furnace at different times. This allows the subsequent feature extraction process to directly extract core data from the reports without having to sift through massive amounts of raw data one by one, significantly improving data processing efficiency.

[0021] Furthermore, In S3, the feature extraction of structured statistical data includes: calculating the breakage rate of the corresponding furnace as a crystallization state feature; calculating the number of breakage warning triggers and the duration of diameter fluctuation standard deviation exceeding the standard in the equal diameter stage within a preset time period as breakage risk features; and calculating the average rate of change of resistivity, oxygen content, and minority carrier lifetime of the corresponding furnace within multiple consecutive sliding windows as quality index features. The standardized feature data is in the format of furnace number-feature type-feature value.

[0022] By accurately calculating core features such as disconnection rate, number of disconnection warning triggers, and average change rate of quality indicators, invalid information is stripped from structured statistical data, and key feature indicators that can directly reflect the abnormal status of the furnace are extracted, achieving the "refinement" of data. The unified data format of "furnace number-feature type-feature value" gives the feature data a high degree of standardization and structure, which facilitates the rapid reading, calculation and threshold comparison of subsequent multi-parameter combination judgment algorithms, greatly improving the efficiency and accuracy of anomaly judgment. At the same time, this standardized format is also suitable for the system requirements of parallel judgment of multiple furnaces, avoiding algorithm processing lag caused by inconsistent formats. In practical implementation, the system's feature extraction unit performs targeted processing on the structured statistical reports of each furnace, calculates the current breakage rate of the furnace as a crystallization state feature, counts the number of breakage warning triggers in the past hour and the duration of diameter fluctuation exceeding the standard deviation in the equal diameter stage as breakage risk features, and calculates the average rate of change of resistivity, oxygen content, and minority carrier lifetime in three consecutive sliding windows as quality indicator features. After extraction, standardized feature data such as "Furnace 05 - Breakage Rate - 70%" and "Furnace 05 - Duration of Diameter Fluctuation Exceeding Standard - 2min" are generated. The algorithm judgment unit can directly read this format data for multi-parameter threshold comparison without additional format conversion, realizing efficient parallel processing of feature extraction and anomaly judgment for 126 furnaces.

[0023] Furthermore, In S3, the dynamic threshold library pre-stores basic judgment thresholds for Czochralski single crystal production, including the judgment thresholds for the standard deviation of diameter fluctuation in the equal diameter stage, the rate of change of furnace temperature, the number of times the wire breakage warning is triggered, the duration of diameter fluctuation exceeding the standard, and the average value of the rate of change of quality indicators. The multi-parameter combination judgment algorithm uniformly divides anomalies into two levels: general anomalies and severe anomalies. Based on the comparison results of the characteristic values ​​of crystal formation state, wire breakage risk, and quality deterioration with the corresponding judgment thresholds, it completes the accurate judgment of the anomaly type and level.

[0024] The anomaly detection process now features quantified and unified reference standards and scientific judgment logic. A dynamic threshold library pre-stores basic thresholds such as the standard deviation of diameter fluctuation during the equal-diameter stage, the rate of temperature change in the furnace, and the number of wire breakage warning triggers. This eliminates the subjectivity of manual judgment, enabling quantitative analysis of anomaly detection. The multi-parameter combination judgment algorithm uniformly classifies anomalies into two levels: general and severe. It combines the comparison results of three characteristic values—crystallization state, wire breakage risk, and quality degradation—with corresponding thresholds to achieve accurate judgment. This overcomes the limitations of traditional single-parameter judgment, effectively identifying anomalies caused by multiple factors and significantly reducing the probability of missed or false alarms. The unified anomaly level classification also provides a clear and actionable basis for subsequent graded warning pushes. In actual implementation, the current wire breakage rate of the furnace is x = QUOTE. The dynamic threshold library presets basic thresholds such as the standard deviation of diameter fluctuation σ≤0.5mm in the equal diameter stage, the furnace temperature change rate k≤±2℃ / min, and the number of wire breakage warning triggers N≤2 times in the past 1 hour. The algorithm judgment unit uses multi-parameter combination judgment logic to carry out anomaly identification. For example, in crystallization anomalies, a wire breakage rate of 50%≤x≤80% is judged as a general crystallization anomaly, and a wire breakage rate x>80% is judged as a severe crystallization anomaly. In wire breakage risk, "N>4 times and σ>1.0mm" is a severe wire breakage risk, and the rest are general wire breakage risks. The algorithm can simultaneously consider the superposition of multiple feature values. For example, when the number of wire breakage warnings exceeds the standard and the diameter fluctuation is too large at the same time, it can accurately judge it as a wire breakage risk anomaly, rather than a one-sided judgment based on a single parameter anomaly. It can still ensure the accuracy and efficiency of the judgment in the scenario of parallel judgment of 126 furnaces.

[0025] Furthermore, In S4, the early warning push method corresponds one-to-one with the anomaly level: general anomalies trigger a pop-up window on the workstation computer + real-time message push to the enterprise, and the pop-up window and message continue until the responsible person confirms receipt; serious anomalies, in addition to the general anomaly push method, add an audible and visual alarm next to the furnace + SMS push to the responsible person's mobile phone, and the audible and visual alarm continues until the anomaly is cleared.

[0026] In actual implementation, when furnace No. 20 is determined to be a general quality deterioration anomaly, the system only triggers a computer pop-up window and a WeChat message push for the corresponding workstation of that furnace. The pop-up window and message will continue to be displayed until the on-site operator and the process engineer of the shift confirm receipt. When the quality deterioration anomaly of furnace No. 20 is upgraded to a serious anomaly, the system immediately activates the audible and visual alarm device next to the furnace, and sends a text message to the relevant responsible person, based on the original online push method. The audible and visual alarm will continue to sound until the furnace anomaly is resolved. The single text message push ensures that the responsible person receives it immediately, and the push message and pop-up window continue until confirmation of receipt. This makes the push for general anomalies simple and efficient, and the push for serious anomalies highly reminder-oriented, realizing differentiated and accurate push for different levels of anomalies.

[0027] Furthermore, In S4, the person responsible for pushing the notification is matched with the level of the anomaly and the furnace platform: general anomalies are matched with the on-site operator and the process engineer of the corresponding furnace platform; severe anomalies are matched with the on-site operator, the process engineer of the corresponding furnace platform, the workshop technical supervisor, and the production administrator.

[0028] The system enables precise and targeted delivery of early warning information. Based on both furnace affiliation and anomaly level, it matches the responsible parties for each warning. This ensures that general anomalies are only sent to the on-site operators and process engineers directly responsible for that furnace, reducing information interference from unrelated personnel and improving the accuracy and efficiency of anomaly response. For serious anomalies, the system adds workshop technical supervisors and production administrators to the basic notification list, allowing high-risk anomalies to receive professional technical guidance and overall coordination at the production management level. This ensures the professionalism and efficiency of anomaly handling and avoids delays or inadequate measures due to improper matching of responsible parties. It truly achieves "responsibility assigned to individuals and tiered handling" in anomaly handling. In practice, furnace No. 45 belongs to shift C. When a general crystallization anomaly occurs, the system only accurately pushes the warning information to the on-site operator and process engineer of shift C responsible for furnace No. 45. The two of them can quickly handle the basic anomaly without the need for other management personnel to intervene. When the crystallization anomaly of the furnace escalates into a serious anomaly, the system will push the warning to the workshop technical supervisor and production administrator in addition to the original push recipients. The workshop technical supervisor provides professional technical guidance for on-site process adjustments, and the production administrator coordinates production resources to ensure that serious anomalies are handled in a timely and professional manner, avoiding the anomaly from escalating and causing greater production losses due to relying solely on on-site operators to handle it.

[0029] Furthermore, In S4, the suggested adjustment direction is generated based on the pre-stored mapping relationship between abnormality type, abnormality feature, and process adjustment scheme. According to the specific type and level of crystallization abnormality, wire breakage risk, and quality deterioration, the corresponding crystal rod growth and furnace environment adjustment strategies are matched.

[0030] Based on a pre-defined mapping relationship between "abnormality type - abnormality characteristics - process adjustment plan", suggested adjustment directions are generated. This ensures that the process adjustment suggestions are not generalized guidance, but rather targeted strategies that are precisely matched to specific abnormality types and levels. After receiving the warning information, the responsible person can directly adjust the process parameters according to the suggestions without having to spend time analyzing and investigating the adjustment direction. This significantly shortens the decision-making time for process adjustments and improves the efficiency of abnormality handling. At the same time, the unified mapping relationship ensures that the process adjustments for the same type of abnormality have standardized characteristics, avoiding large differences in the adjustment methods when different responsible persons handle the same type of abnormality, and ensuring the consistency of the Czochralski single crystal production process. In actual implementation, the system's targeted real-time early warning push module has a built-in database of precise mapping relationships between anomaly types and process adjustment suggestions. When furnace 18 is determined to be a general crystallization anomaly (68% breakage rate), the system automatically matches the corresponding suggested adjustment direction: fine-tune the crucible ratio (±0.05) and reduce the crystal rod growth rate (0.1mm / min). When the crystallization anomaly of furnace 18 is upgraded to a severe anomaly (86% breakage rate), the system matches the adjustment suggestions corresponding to the severe anomaly: increase the argon flow rate (±10sl / min), switch the induction and release to equal crucible rotation (+1r / min), and increase the furnace pressure by 0.2kPa. On-site operators can directly and quickly adjust process parameters according to these suggestions without additional process analysis, which greatly improves the speed and pertinence of process adjustments. Moreover, the adjustment suggestions for handling similar anomalies by shifts A, B, and C are completely consistent, ensuring the consistency of the entire plant's production process.

[0031] Furthermore, In S6, the specific rules for closed-loop determination are as follows: if all real-time feature data after process adjustment within a preset time period returns to the judgment threshold range of the dynamic threshold library, the closed loop is determined to be successful, the corresponding furnace platform is marked as resolved and the normal data acquisition frequency is restored; if the feature data does not all return to the threshold range within a preset time period, the adjustment is determined to be invalid, the abnormality level is upgraded and a second early warning push is triggered, and continuous enhanced monitoring is carried out until the abnormality is resolved.

[0032] A complete closed-loop control system of "monitoring-early warning-adjustment-verification" has been constructed. The core criterion for the success of the closed loop is whether all characteristic data after process adjustment returns to the threshold range of the dynamic threshold library. This provides an objective and unified quantitative standard for verifying the effect of process adjustment, avoiding the subjectivity and arbitrariness of manual verification. The setting of restoring the normal data collection frequency after successful closed loop can effectively save the system's computing and communication resources and adapt to the needs of large-scale monitoring of multiple furnaces. The escalation of the anomaly level and the push of secondary early warning when the adjustment is ineffective ensure that the anomaly is continuously monitored by the system, preventing the anomaly from further expanding after the process adjustment is ineffective. At the same time, continuous post-enhanced monitoring can track the furnace status in real time, providing accurate and continuous data analysis support for subsequent secondary handling. Ultimately, the handling of anomalies forms a complete feedback chain, realizing the continuous optimization of the production process. In actual implementation, when furnace 33 triggers an early warning due to a general disconnection risk, the responsible person completes the process adjustment as suggested by the system. The system immediately initiates enhanced post-monitoring of the furnace, increasing the data acquisition frequency from the usual 5 minutes / time to 2.5 minutes / time, continuously collecting characteristic data for 30 minutes. If, within 30 minutes, all core characteristic data such as the number of disconnection warning triggers and the standard deviation of diameter fluctuation in the equal diameter stage return to the threshold range of the dynamic threshold library, the system automatically determines that the closed loop is successful, marks the furnace 33 as having resolved its anomaly, and restores the acquisition frequency to 5 minutes / time. If, within 30 minutes, the characteristic data does not all return to the threshold range, the system determines that the adjustment is ineffective, upgrades the general disconnection risk of the furnace to a severe disconnection risk, triggers a secondary early warning, and continues to collect data at a frequency of 2.5 minutes / time until the anomaly is resolved. This rule provides clear time and quantitative standards for verifying the effectiveness of process adjustments, completely solving the problem of "no verification after adjustment and no closed loop for anomaly handling" in traditional monitoring.

[0033] Furthermore, A real-time online monitoring and early warning system for Czochralski single crystals, used to implement the real-time online monitoring and early warning method for Czochralski single crystals as described above, characterized in that it includes: a multi-dimensional data acquisition module, a furnace-level automatic statistical analysis module, an anomaly intelligent judgment module, a targeted real-time early warning push module, a closed-loop handling and tracking module, and a real-time performance assurance module; The multi-dimensional data acquisition module establishes a gigabit network high-speed communication connection with the industrial control system of each Czochralski single crystal furnace to collect multi-dimensional raw data and complete standardized naming, and output standardized named raw data. The furnace-level automatic statistical analysis module is connected to the multi-dimensional data acquisition module to receive standardized named raw data, classify it according to the furnace number, and generate structured statistical data through sliding window statistical analysis. The anomaly intelligent judgment module is communicatively connected to the furnace-level automatic statistical analysis module. It has a built-in dynamic threshold library for extracting features from structured statistical data to generate standardized feature data. It also completes anomaly judgment through a multi-parameter combination judgment algorithm and outputs standardized anomaly judgment results. The targeted real-time early warning push module is communicatively connected to the anomaly intelligent judgment module. It is used to complete the hierarchical early warning triggering and push responsibility matching based on the standardized anomaly judgment results, and push early warning information to the corresponding responsible person's terminal. The closed-loop handling tracking module is connected to the directional real-time early warning push module and the anomaly intelligent judgment module. It has a built-in knowledge base and is used to start the furnace platform post-enhanced monitoring, collect real-time feature data after process adjustment, verify the adjustment effect and output the closed-loop judgment result. When the closed loop is successful, it stores the abnormal cases and updates and optimizes the dynamic threshold library, and feeds the optimized dynamic threshold library back to the anomaly intelligent judgment module. The real-time performance assurance module establishes communication connections with all other modules to provide real-time support for the entire system through edge computing, gigabit network high-speed communication, and multi-threaded parallel processing technology, ensuring that the total latency of the entire process from data acquisition to early warning push meets the requirements of real-time monitoring of Czochralski single crystal.

Claims

1. A method for real-time online monitoring and early warning of Czochralski single crystals, characterized in that, Includes the following steps: S1. Collect raw data on crystal formation status, wire breakage, and quality inspection of each furnace in real time, and standardize and name them, then output the raw data with standardized names. S2. Furnace-level data statistics and standardization processing: After the raw data is classified by furnace number according to the standardized naming, it is statistically analyzed by sliding window to generate structured statistical data. S3. Extract features from structured statistical data to generate standardized feature data, retrieve the dynamic threshold library and use a multi-parameter combination judgment algorithm to complete the anomaly type identification and level classification, and output the standardized anomaly judgment result. S4. Based on the standardized anomaly judgment results, trigger the early warning push method according to the anomaly level, match the push responsibility according to the furnace and anomaly type, and push early warning information including suggested adjustment directions; S5. Enhanced monitoring after process adjustment: After the responsible person confirms receipt of the pushed early warning information, the enhanced monitoring of the corresponding furnace is started, the collection frequency is increased to continuously collect the characteristic data after process adjustment, and the real-time characteristic data after process adjustment is output. S6. Verification of Adjustment Effect and Closed-Loop Judgment: The effectiveness of the process adjustment is verified by comparing the feature data after the process adjustment with the judgment threshold range of the normal range of feature data in the dynamic threshold library. If the feature data is within the normal feature data range, the closed-loop judgment result is output; otherwise, the closed-loop judgment result of the adjustment is invalid is output.

2. The method for real-time online monitoring and early warning of Czochralski single crystal according to claim 1, characterized in that, In S1, the raw data on crystal formation status includes the crystal rod diameter, growth rate, furnace temperature, and argon flow rate at each process stage of the Czochralski single crystal furnace, including crystal introduction, shoulder formation, equal diameter formation, and finishing. The raw data related to wire breakage includes the historical number of wire breaks in the Czochralski single crystal furnace, the process stage in which the wire breakage occurred, the fluctuation of the crystal rod diameter within a preset time before the wire breakage, and the fluctuation data of the furnace temperature. The raw data for quality inspection includes the resistivity, oxygen content, and minority carrier lifetime data of the crystal rod.

3. The method for real-time online monitoring and early warning of Czochralski single crystal according to claim 1, characterized in that, In S2, the statistical analysis of the sliding window includes calculating the average value of crystallization state data, the frequency of wire breakage related data, and the change rate of quality inspection data within the window. The structured statistical data includes the crystallization state trend curve of the corresponding furnace, the wire breakage risk frequency report, and the quality index change report.

4. The method for real-time online monitoring and early warning of Czochralski single crystal according to claim 1, characterized in that, In S3, the feature extraction of structured statistical data includes: calculating the breakage rate of the corresponding furnace as a crystallization state feature; calculating the number of breakage warning triggers and the duration of diameter fluctuation standard deviation exceeding the standard in the equal diameter stage within a preset time period as breakage risk features; and calculating the average rate of change of resistivity, oxygen content, and minority carrier lifetime of the corresponding furnace within multiple consecutive sliding windows as quality index features. The standardized feature data is in the format of furnace number-feature type-feature value.

5. The method for real-time online monitoring and early warning of Czochralski single crystal according to claim 4, characterized in that, In S3, the dynamic threshold library pre-stores basic judgment thresholds for Czochralski single crystal production, including the judgment thresholds for the standard deviation of diameter fluctuation in the equal diameter stage, the rate of change of furnace temperature, the number of times the wire breakage warning is triggered, the duration of diameter fluctuation exceeding the standard, and the average value of the rate of change of quality indicators. The multi-parameter combination judgment algorithm uniformly divides anomalies into two levels: general anomalies and severe anomalies. Based on the comparison results of the characteristic values ​​of crystal formation state, wire breakage risk, and quality deterioration with the corresponding judgment thresholds, it completes the accurate judgment of the anomaly type and level.

6. The method for real-time online monitoring and early warning of Czochralski single crystal according to claim 5, characterized in that, In S4, the early warning push method corresponds one-to-one with the anomaly level: general anomalies trigger a pop-up window on the workstation computer + real-time message push to the enterprise, and the pop-up window and message continue until the responsible person confirms receipt; serious anomalies, in addition to the general anomaly push method, add an audible and visual alarm next to the furnace + SMS push to the responsible person's mobile phone, and the audible and visual alarm continues until the anomaly is cleared.

7. The method for real-time online monitoring and early warning of Czochralski single crystal according to claim 5, characterized in that, In S4, the person responsible for pushing the notification is matched with the level of the anomaly and the furnace platform: general anomalies are matched with the on-site operator and the process engineer of the corresponding furnace platform; severe anomalies are matched with the on-site operator, the process engineer of the corresponding furnace platform, the workshop technical supervisor, and the production administrator.

8. The method for real-time online monitoring and early warning of Czochralski single crystal according to claim 5, characterized in that, In S4, the suggested adjustment direction is generated based on the pre-stored mapping relationship between abnormality type, abnormality feature, and process adjustment scheme. According to the specific type and level of crystallization abnormality, wire breakage risk, and quality deterioration, the corresponding crystal rod growth and furnace environment adjustment strategies are matched.

9. The method for real-time online monitoring and early warning of Czochralski single crystal according to claim 1, characterized in that, In S6, the specific rules for closed-loop determination are as follows: if all real-time feature data after process adjustment within a preset time period return to the determination threshold range of the dynamic threshold library, then the closed loop is determined to be successful, the corresponding furnace platform is marked as abnormal and the normal data acquisition frequency is restored. If the feature data does not all return to the threshold range within the preset time period, the adjustment is deemed invalid, the anomaly level is upgraded and a second warning is triggered, and continuous enhanced monitoring continues until the anomaly is resolved.

10. A real-time online monitoring and early warning system for Czochralski single crystals, used to implement the real-time online monitoring and early warning method for Czochralski single crystals as described in any one of claims 1-9, characterized in that, It includes a multi-dimensional data acquisition module, a furnace-level automatic statistical analysis module, an anomaly intelligent judgment module, a targeted real-time early warning push module, a closed-loop handling and tracking module, and a real-time assurance module; The multi-dimensional data acquisition module establishes a gigabit network high-speed communication connection with the industrial control system of each Czochralski single crystal furnace to collect multi-dimensional raw data and complete standardized naming, and output standardized named raw data. The furnace-level automatic statistical analysis module is connected to the multi-dimensional data acquisition module to receive standardized named raw data, classify it according to the furnace number, and generate structured statistical data through sliding window statistical analysis. The anomaly intelligent judgment module is communicatively connected to the furnace-level automatic statistical analysis module. It has a built-in dynamic threshold library for extracting features from structured statistical data to generate standardized feature data. It also completes anomaly judgment through a multi-parameter combination judgment algorithm and outputs standardized anomaly judgment results. The targeted real-time early warning push module is communicatively connected to the anomaly intelligent judgment module. It is used to complete the hierarchical early warning triggering and push responsibility matching based on the standardized anomaly judgment results, and push early warning information to the corresponding responsible person's terminal. The closed-loop handling tracking module is connected to the directional real-time early warning push module and the anomaly intelligent judgment module, respectively. It has a built-in knowledge base and is used to start the furnace platform post-enhanced monitoring, collect real-time feature data after process adjustment, verify the adjustment effect and output the closed-loop judgment result. When the closed loop is successful, it stores the abnormal cases and updates and optimizes the dynamic threshold library, and feeds the optimized dynamic threshold library back to the anomaly intelligent judgment module. The real-time performance assurance module establishes communication connections with all other modules to provide real-time support for the entire system through edge computing, gigabit network high-speed communication, and multi-threaded parallel processing technology, ensuring that the total latency of the entire process from data acquisition to early warning push meets the requirements of real-time monitoring of Czochralski single crystal.

Images