Control method of single crystal furnace, electronic device, storage medium, and program product

Abstract

Embodiments of the present application provide a control method of a single crystal furnace, an electronic device, a storage medium and a program product. The method comprises: collecting real-time running data of a single crystal furnace, and extracting key feature waveforms of an abnormal period of the single crystal furnace; inputting the real-time running data into a preset diagnosis model for abnormality identification; when an abnormal event is identified, performing root cause tracing reasoning through a knowledge graph to locate an abnormal root cause; according to the abnormal root cause, retrieving matched historical measures and / or standard operating procedures from a measure library to generate candidate instructions containing specific parameter adjustment instructions; and simulating and verifying the candidate instructions in a digital simulation model, and if the simulation is passed, executing the candidate instructions, wherein the digital simulation model is used to simulate the running state of the single crystal furnace.

Description

Technical Field

[0001] This application relates to the field of single-crystal silicon preparation technology, and in particular to a control method for a single-crystal furnace, electronic equipment, storage medium, and program product. Background Technology

[0002] Monocrystalline silicon is a core material for semiconductor chips and photovoltaic cells. Currently, the mainstream method for growing monocrystalline silicon is the Czochralski method (CZ method) in a single crystal furnace. During the growth of monocrystalline silicon, multiple process parameters, such as pulling speed, temperature, crystal diameter, crucible height ratio, and argon flow rate, need to be controlled in real time to ensure crystal quality and avoid abnormal events such as wire breakage, dislocations, and excessive oxygen content.

[0003] Currently, production line operations primarily rely on the manual experience of process engineers for anomaly diagnosis and process adjustments. Specifically, when the monitoring system indicates an anomaly or operators discover phenomena such as crystal diameter fluctuations or temperature drift, engineers retrieve historical curves recorded by the data acquisition system, combine this with their own experience to determine the cause of the anomaly, and then manually modify the process settings in the programmable logic controller or manufacturing execution system in an attempt to restore stable growth.

[0004] Current methods for handling anomalies in monocrystalline silicon production rely on human experience, resulting in delayed anomaly responses, highly subjective diagnostic results, and difficulty in passing on tacit knowledge. There is an urgent need for an intelligent solution that can achieve automatic anomaly diagnosis and real-time closed-loop control. Summary of the Invention

[0005] This application provides a control method for a single crystal furnace to solve the aforementioned technical problems. This method addresses the full-process control needs of anomaly identification, root cause localization, and handling decisions in single crystal silicon production. It comprehensively analyzes multi-source data during the operation of the single crystal furnace, accurately locates the root cause of anomalies based on the identification of abnormal events, and forms executable control decisions by combining historical handling experience and standard operating procedure information. Furthermore, the control decisions are pre-verified before execution, thereby improving the timeliness, accuracy, and stability of anomaly handling in the single crystal furnace.

[0006] In a first aspect, embodiments of this application provide a method for controlling a single crystal furnace, including:

[0007] Collect real-time operating data of the single crystal furnace and extract key characteristic waveforms during abnormal periods of the single crystal furnace;

[0008] Real-time operational data is input into a preset diagnostic model for anomaly identification. When an abnormal event is identified, the root cause is located by using a knowledge graph for root cause reasoning.

[0009] Based on the root cause of the anomaly, retrieve matching historical measures and / or standard operating procedures from the measures library to generate candidate instructions containing specific parameter adjustment instructions;

[0010] Candidate instructions are simulated and verified in the digital simulation model. If the simulation passes, the candidate instructions are issued for execution. The digital simulation model is used to simulate the operating state of the single crystal furnace.

[0011] In one possible embodiment,

[0012] Real-time operational data is input into a pre-defined diagnostic model for anomaly identification. When an abnormal event is identified, root cause reasoning is performed using a knowledge graph to pinpoint the root cause of the anomaly, including:

[0013] Real-time running data is input into a preset diagnostic model, which includes a mechanism diagnostic sub-model and a data-driven diagnostic sub-model that run in parallel.

[0014] The mechanism diagnosis sub-model is constructed based on the thermal field mechanics equation or the fluid mechanics equation, and is used to determine whether the current operating state deviates from the theoretical expectation from the perspective of physical mechanism, and output the first anomaly probability.

[0015] The data-driven diagnostic sub-model is a deep neural network model trained based on historical data of wire breakage accidents or quality anomalies. It is used to identify abnormal patterns from a data statistics perspective and output a second anomaly probability.

[0016] When at least one of the first abnormal probability and the second abnormal probability exceeds a preset alarm threshold, it is determined that an abnormal event exists and the abnormal type of the abnormal event is determined.

[0017] Starting with the anomaly type, a reverse reasoning algorithm is used to perform multi-hop association queries in the knowledge graph. The knowledge graph contains at least equipment nodes, process parameter nodes, raw material attribute nodes, and the relationships between the three.

[0018] By tracing back layer by layer along the association relationship, the conditional probability corresponding to each candidate root cause node is calculated, and the root cause node with the highest conditional probability is selected as the abnormal root cause.

[0019] In one possible embodiment, simulation verification of candidate instructions is performed in a digital simulation model, including:

[0020] The candidate instructions are input into the digital simulation model to simulate the parameter adjustment actions corresponding to the candidate instructions. The digital simulation model is used to simulate the predicted operating state of the single crystal furnace under the candidate instructions.

[0021] Obtain at least one predicted operating index output by the digital simulation model after the simulation is executed, and determine whether the predicted operating index meets the preset safety and stability conditions.

[0022] If the safety and stability conditions are met, the simulation is deemed successful, and the candidate instructions can be issued for execution.

[0023] If the safety and stability conditions are not met, the simulation is deemed unsuccessful, the candidate instructions are abandoned, and alternative instructions are retrieved from the measure library for re-simulation verification.

[0024] In one possible embodiment, if the safety and stability conditions are not met, the simulation is deemed unsuccessful, the candidate instructions are discarded, and alternative instructions are retrieved from the measure library for re-simulation verification. The method further includes:

[0025] Multiple alternative instructions are retrieved from the measure library, and the predicted operating indicators of each alternative instruction after execution in the digital simulation model are estimated.

[0026] Based on the overall deviation between each predicted operating indicator and its corresponding target operating indicator, the candidate instructions are ranked, and the candidate instruction with the smallest overall deviation is selected as the new candidate instruction.

[0027] The new candidate instructions are re-input into the digital simulation model for simulation verification until a candidate instruction that passes the simulation is found or all alternative instructions have been retrieved.

[0028] In one possible embodiment, the candidate instruction is issued for execution, and the method further includes:

[0029] The diagnostic report and simulation verification results are pushed to the engineer through the human-machine interface, and a manual confirmation timer is started;

[0030] If an acknowledgment signal from the engineer is received within a preset time, the candidate instruction is sent to the programmable logic controller or manufacturing execution system via the industrial communication protocol in response to the acknowledgment signal to modify the process settings.

[0031] If no confirmation signal is received within the preset time, the candidate instruction will be automatically sent to the programmable logic controller or manufacturing execution system via the industrial communication protocol.

[0032] In one possible embodiment, after acquiring the real-time operating data of the single crystal furnace, the process further includes:

[0033] Clean, align and extract features from real-time operating data to extract at least one of the following during abnormal periods: the oscillation decay waveform of crystal diameter, the time-delay response waveform of liquid surface temperature or the periodic fluctuation waveform of pulling speed.

[0034] Extract the waveform's morphological feature parameters, which include at least one of the following: oscillation frequency, attenuation coefficient, response delay, or fluctuation amplitude.

[0035] Input morphological feature parameters into a preset diagnostic model to identify abnormal events.

[0036] In one possible embodiment, after the candidate instruction is issued and executed, the method further includes:

[0037] Collect actual running data after execution and record the operation log for this exception handling. The operation log includes: exception type, root cause of exception located, specific parameters of candidate instructions, whether the engineer modified the candidate instructions, content of the modified instructions, and evaluation of the control effect after execution.

[0038] Automatically save operation logs to the knowledge base;

[0039] When the preset triggering conditions are met, the data-driven diagnostic sub-model is incrementally trained using the newly added data in the knowledge base, the entity relationships and attributes of the knowledge graph are updated, and the historical measures in the measures library and their application effect scores are dynamically adjusted.

[0040] Secondly, embodiments of this application provide a control device for a single crystal furnace, comprising:

[0041] The acquisition module is used to collect real-time operating data of the single crystal furnace and extract key characteristic waveforms during abnormal periods of the single crystal furnace.

[0042] The identification module is used to input real-time running data into a preset diagnostic model for anomaly identification. When an abnormal event is identified, the root cause is located by using a knowledge graph for root cause reasoning.

[0043] The matching module is used to retrieve matching historical measures and / or standard operating procedures from the measures library based on the root cause of the anomaly, and generate candidate instructions containing specific parameter adjustment instructions;

[0044] The execution module is used to simulate and verify candidate instructions in the digital simulation model. If the simulation passes, the candidate instructions are issued for execution. The digital simulation model is used to simulate the operating state of the single crystal furnace.

[0045] Thirdly, embodiments of this application provide a control device for a single crystal furnace, including: a memory and a processor;

[0046] The memory stores computer-executed instructions;

[0047] The processor executes computer execution instructions stored in the memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as described above.

[0048] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect and / or various possible implementations of the first aspect.

[0049] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the first aspect and / or various possible implementations of the first aspect.

[0050] The control method, electronic device, storage medium, and program product for a single crystal furnace provided in this application improve the timeliness of anomaly identification and the accuracy of anomaly root cause location by collecting real-time operating data of the single crystal furnace, extracting key characteristic waveforms during abnormal periods, combining anomaly identification with a preset diagnostic model, and using knowledge graphs for root cause reasoning. By retrieving matching historical measures and / or standard operating procedures from the measures library based on the anomaly root cause, generating candidate instructions containing specific parameter adjustment instructions, and simulating and verifying the candidate instructions in a digital simulation model before issuing them for execution, the pertinence and stability of anomaly handling decisions can be improved, thereby reducing reliance on human experience and enhancing the reliability of single crystal furnace operation control. Attached Figure Description

[0051] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0052] Figure 1 Flowchart of the control method for the single crystal furnace provided in this application Figure 1 ;

[0053] Figure 2 Flowchart of the control method for the single crystal furnace provided in this application Figure 2 ;

[0054] Figure 3 A schematic diagram of the control device for the single crystal furnace provided in this application;

[0055] Figure 4 This is a schematic diagram of the control equipment for the single crystal furnace provided in this application.

[0056] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0057] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0058] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0059] The Czochralski (CZ) production of monocrystalline silicon is a critical process in semiconductor and photovoltaic manufacturing, primarily used in continuous control scenarios involving the melting, nucleation, crystal pulling, and crystal growth of high-purity polycrystalline silicon raw materials in a single-crystal furnace. In these scenarios, the single-crystal furnace typically operates in conjunction with a data acquisition system, process monitoring system, and equipment control unit, monitoring and adjusting key parameters such as temperature field, pulling speed, gas flow rate, furnace pressure, and liquid level in real time. Because the monocrystalline silicon growth process demands extremely high thermal stability and consistency of process parameters, even minor fluctuations can quickly escalate into serious anomalies such as crystal diameter instability, increased dislocation defects, abnormal oxygen and carbon content, or even broken crystals. Therefore, continuous data observation and rapid response mechanisms are essential to ensure production stability. In actual production, process engineers often need to combine historical curves, alarm information, and equipment status to assess the operating trend of the single-crystal furnace and take timely corrective measures when anomalies occur to prevent the escalation of anomalies from affecting the overall product quality and equipment safety. Such application scenarios place high demands on diagnostic timeliness, analytical accuracy, and handling stability, which also determines the important role of the anomaly identification and handling system in monocrystalline silicon production.

[0060] Current methods for handling anomalies in monocrystalline silicon furnaces typically rely on manual experience analysis and threshold-based monitoring alarms. The operational logic usually involves collecting real-time data, displaying parameter curves such as temperature, pulling speed, and gas flow rate on a monitoring interface, and setting corresponding thresholds or fluctuation ranges. When a parameter exceeds the preset range, an alarm is triggered. Engineers then combine historical data and process experience to determine the nature and cause of the anomaly and manually determine subsequent adjustments. This approach can provide basic early warning when facing single parameter exceedances or clearly defined anomalies. However, in monocrystalline silicon production, many anomalies are not triggered by a single variable but rather evolve from the coupling of multiple factors such as changes in thermal field distribution, atmosphere disturbances, pulling speed fluctuations, and changes in raw material state. Furthermore, the anomaly trend is often gradual and insidious; by the time the threshold is triggered, the anomaly's impact has often accumulated to a significant degree. Because current methods primarily focus on whether a single parameter exceeds its limits, they cannot perform correlation analysis on multi-source data such as temperature, pulling speed, and gas flow rate, nor can they identify the anomaly evolution chain from the dynamic relationships between parameters. Therefore, root cause localization often relies on personal experience and manual investigation, resulting in low diagnostic efficiency and poor consistency. For similar anomalies, different engineers may arrive at different conclusions, leading to inconsistent handling strategies. Furthermore, manual analysis of historical curves is typically time-consuming and difficult to respond to production cycles promptly. More critically, threshold alarms only indicate an anomaly but do not provide actionable solutions. Engineers still need to rely on experience to select adjustment methods, which can easily result in inappropriate adjustments, directional deviations, or delayed responses, potentially triggering secondary anomalies and impacting crystal quality and continuous equipment operation. Therefore, existing technologies have significant shortcomings in terms of the timeliness of anomaly identification, the accuracy of root cause localization, and the stability of subsequent handling, making it difficult to meet the requirements of monocrystalline silicon Czochralski production for a highly reliable, strongly correlated, and fast-response anomaly management model.

[0061] In light of this, how to achieve rapid identification of abnormal periods, accurate root cause localization, and more stable handling during the Czochralski production process of monocrystalline silicon has become an urgent technical problem to be solved. To address this issue, a technical concept for monocrystalline furnace control has been developed. This involves collecting real-time operating data from the monocrystalline furnace and extracting key characteristic waveforms during abnormal periods. The real-time operating data is then input into a pre-defined diagnostic model for anomaly identification. Once an abnormal event is identified, root cause reasoning is further conducted using a knowledge graph to locate the root cause. Subsequently, based on the root cause, matching historical measures and / or standard operating procedures are retrieved from a measure library to generate candidate instructions containing specific parameter adjustment commands. Finally, the candidate instructions are simulated and verified in a digital simulation model. If the verification is successful, the candidate instructions are issued for execution. The digital simulation model is used to simulate the operating state of the monocrystalline furnace. This complete technical route, from data acquisition, anomaly identification, root cause reasoning, measure generation to simulation verification and execution, can transform anomaly handling from simple ex-post experience-based judgment to an orderly processing flow oriented towards operational data and knowledge association, without changing the basic control object and production scenario of the single crystal furnace. This provides a clear technical path foundation for the subsequent development of diagnosis, reasoning, measure selection and simulation verification in specific implementation methods.

[0062] Figure 1 Flowchart of the control method for the single crystal furnace provided in this application Figure 1 ,like Figure 1 As shown, the method includes:

[0063] S101. Collect real-time operating data of the single crystal furnace and extract key characteristic waveforms during abnormal periods of the single crystal furnace.

[0064] In this embodiment, the single crystal furnace can be a set of control equipment for the Czochralski production of single crystal silicon, including at least a heating system, a crucible lifting mechanism, a seed crystal lifting mechanism, a gas supply system, a furnace pressure control unit, a temperature detection unit, and a controller connected thereto. Real-time operating data refers to the set of process parameters, equipment status parameters, and alarm status parameters continuously acquired according to a set sampling period during the operation of the single crystal furnace. Specifically, it can include time-series data such as pulling speed, crystal diameter, heater power, crucible-to-lift ratio, argon flow rate, furnace pressure, liquid level, thermal field temperature, cooling water status, actuator feedback values, and equipment interlock status. Key characteristic waveforms refer to time-series waveform segments and their derived features that can characterize abnormal evolution trends or abnormal formation mechanisms. These can be original parameter curves or waveform expressions obtained through filtering, transformation, fitting, and feature extraction, such as the oscillation frequency of crystal diameter fluctuations, the slope of the temperature drift curve, the transition amplitude of sudden changes in gas flow rate, the periodic disturbance intensity of the pulling speed waveform, and the phase difference between multiple parameters.

[0065] In practical implementation, data acquisition can be completed by an on-site monitoring and data acquisition system. This system can connect to a PLC (Programmable Logic Controller), temperature controller, flow controller, frequency converter, sensor acquisition module, and MES (Manufacturing Execution System) via industrial Ethernet, fieldbus, or serial communication interface. This allows for continuous acquisition of multi-source heterogeneous data during different process stages of the single crystal furnace, such as melting, crystal pulling, necking, shoulder formation, constant diameter growth, and finishing. To ensure comparability between different data sources, the raw data is first processed after acquisition, including timestamp unification, abnormal null value removal, duplicate sampling deduplication, and noise suppression. Noise suppression can be achieved using moving average, median filtering, or Kalman filtering. For temperature, flow, and position feedback data with different sampling frequencies, they can be uniformly resampled to the same time base. For data channels with time delay offsets, a dynamic time warping algorithm can be introduced for waveform alignment to eliminate misalignment caused by control command response lag and sampling link delay. After preprocessing, the system identifies suspected abnormal periods based on historical normal operation templates and real-time window statistics. Suspected abnormal periods can be triggered by conditions such as a sudden increase in variance within the sliding window, abrupt changes in slope, abnormal spectral energy, or a decrease in multivariate correlation. For identified abnormal periods, key characteristic waveforms are further extracted. Extraction methods include extracting oscillation segments from the crystal diameter curve and calculating the dominant frequency and attenuation coefficient; extracting drift trends and local fluctuation envelopes from the temperature curve; extracting the coupling change amplitude from the argon flow rate and furnace pressure linkage curve; and extracting overshoot and recovery time from the pulling speed and diameter feedback curve. This process forms a unified abnormal feature data package, which, along with the original operating data for the corresponding period, is sent to the subsequent diagnostic stage.

[0066] Based on the above analysis, this step, through continuous acquisition, cleaning, alignment, and waveform feature extraction of multi-source operating parameters of the single-crystal furnace, expands the traditional monitoring method that relies solely on single-parameter limit exceedances into an anomaly characterization method oriented towards time-series correlations. Since anomalies in the Czochralski process of single-crystal silicon are typically gradual and coupled, relying solely on whether a single parameter exceeds its limit is insufficient to reflect the anomaly formation process in a timely manner. Key feature waveforms, however, can preserve the trend, rhythm, and linkage of parameter changes, thus providing a more complete input basis for subsequent anomaly identification. Especially before problems such as crystal diameter instability, thermal field drift, and atmosphere disturbances develop into obvious alarms, waveform-level feature extraction can capture early signs of anomalies, thereby shortening the anomaly detection time and reducing the search range for subsequent root cause localization.

[0067] S102. Input real-time running data into the preset diagnostic model for anomaly identification. When an abnormal event is identified, use the knowledge graph to perform root cause reasoning and locate the root cause of the anomaly.

[0068] In this embodiment, a pre-defined diagnostic model is used to classify, detect, or issue early warnings for the operating status of the single crystal furnace. It can output not only the judgment of whether an anomaly exists, but also the anomaly type, probability, severity, and development trend. An abnormal event refers to an event determined by the diagnostic model that characterizes a deviation of the single crystal furnace from the target process state and requires further analysis and handling. Examples include abnormal diameter oscillation, abnormal thermal field imbalance, abnormal wire breakage risk, abnormal atmosphere fluctuation, abnormal liquid level, and abnormal actuator response. A knowledge graph is a structured knowledge network used to describe the relationships between single crystal furnace anomalies, process parameters, equipment components, physical mechanisms, historical cases, and handling measures. Nodes can represent abnormal phenomena, equipment states, process stages, controlled objects, root causes, or measures, while edges can represent causal relationships, dependencies, temporal sequences, constraints, or correlation probabilities. Root cause reasoning refers to the process of performing association retrieval, evidence aggregation, and confidence calculation along causal paths in the knowledge graph based on identified abnormal events and their contextual parameter states to determine the root cause of the target anomaly from multiple candidate causes.

[0069] In practical implementation, the pre-defined diagnostic model can adopt a hybrid architecture that combines a mechanistic model and a data-driven model. The mechanistic model is established based on physical laws such as heat conduction, convection heat transfer, melt flow, crystal growth interface stability, and actuator dynamic response in a single crystal furnace. It can provide interpretable deviation judgments for issues such as thermal field shifts, liquid level changes, and control hysteresis. The data-driven model is trained using historical production data, alarm data, process logs, and fault records. The model can be in the form of a deep neural network, temporal convolutional network, long short-term memory network, Transformer temporal model, or ensemble learning classifier, used to identify complex coupled anomaly patterns. After real-time data input, the mechanistic model generates a physical consistency deviation index, and the data-driven model generates anomaly classification probability vectors. When the output of either model reaches a pre-defined trigger condition, the system determines that an anomaly event has occurred and initiates the root cause tracing process. The trigger condition can be set as an anomaly probability exceeding a threshold, multiple consecutive windows determining anomalies, or physical residuals exceeding the process allowable range. Subsequently, the system maps the anomaly type, process stage, key characteristic waveform parameters, equipment status, and upstream and downstream variables together as knowledge graph query conditions. For example, when an anomaly indicating a risk of line breakage is identified, the system first locates upstream nodes associated with the risk, such as fluctuations in liquid surface temperature, abnormal melt convection, abnormal argon flow, pulling speed disturbances, and heater power fluctuations. It then further extends upstream to candidate root cause nodes such as valve jamming, controller setpoint drift, aging of thermal components, or changes in raw material condition. For each candidate path, the system comprehensively calculates the confidence level by combining historical statistical probabilities, the degree of matching of current observational evidence, and path length. Formally, this can be expressed as the root cause score equal to the prior association weight multiplied by the current evidence matching coefficient and then multiplied by the temporal consistency factor. The prior association weight reflects the strength of the association between the anomaly and the root cause in historical cases, the current evidence matching coefficient reflects the similarity between real-time data and the typical characteristics of the root cause, and the temporal consistency factor reflects whether the order of changes in each variable conforms to the evolutionary mechanism of the root cause. After sorting, the most probable anomaly root cause is obtained, and the root cause identifier, confidence level, associated evidence, and reasoning path are output to the measure generation stage. In one possible embodiment, the knowledge graph query path can also be dynamically adjusted by a reinforcement learning strategy, such as using a deep Q (Action-Value) network to select the next-hop query direction based on past reasoning success rates and current evidence distribution, thereby reducing search costs when there are many candidate root causes.

[0070] Based on the above analysis, this step, by inputting real-time data into the diagnostic model and combining it with a knowledge graph to perform root cause analysis, elevates anomaly management from simple limit-crossing alarms to a dual judgment mechanism based on data patterns and knowledge relationships. The mechanistic model ensures the diagnostic process aligns with actual physical laws, the data-driven model captures complex coupling patterns that are difficult to explicitly model manually, and the knowledge graph connects anomalies causally with equipment, processes, and historical experience, thus improving the timeliness of anomaly identification and the consistency of root cause localization. Compared to relying on manual curve analysis for item-by-item troubleshooting, this step transforms the dynamic relationships between multiple source parameters into computable reasoning paths, reducing judgment biases caused by differences in experience among engineers and providing a clear basis for generating targeted handling instructions.

[0071] S103. Based on the root cause of the anomaly, retrieve matching historical measures and / or standard operating procedures from the measures library, and generate candidate instructions containing specific parameter adjustment instructions.

[0072] In this embodiment, the measure library is a structured set of handling knowledge established for various abnormal root causes of single crystal furnaces. It includes at least historical successful handling cases, standard operating procedures, manual intervention templates confirmed by process engineers, execution constraints, applicable process stages, dependencies between measures, and post-execution effect evaluation data. Historical measures refer to adjustment plans implemented and recorded for specific abnormal root causes during historical production processes. Standard operating procedures refer to standardized handling processes pre-defined for typical abnormalities. Candidate instructions refer to a set of control instructions to be verified, generated based on the current abnormal root cause, current operating conditions, and measure adaptation results. These control instructions include the specific adjustment object, adjustment direction, adjustment magnitude, adjustment rate, execution sequence, and holding duration, and can be identified and executed by the control system or operating terminal.

[0073] In practice, after obtaining the root cause of the anomaly, the system first searches the measure library based on the root cause identifier, process stage, current diameter specification, equipment model, raw material batch, and constraint boundaries. The search can employ either rule-based matching or vector similarity recall. In rule-based matching, an index relationship of "root cause—process stage—equipment status—measure template" can be established. For example, when the root cause is abnormal argon flow and the system is in the constant diameter growth stage, standard treatment templates related to atmosphere stabilization and recovery are prioritized for retrieval; when the root cause is heater power drift leading to thermal field shift, historical measures related to power correction and crucible rise coordination are prioritized for retrieval. In vector similarity recall, the current anomaly context can be encoded into a feature vector, compared with historical case vectors to recall several similar cases, and extract the measure combinations with higher success rates. Based on the search results, the system performs adaptation screening, which includes safety constraint checks, actuator capability checks, and process boundary checks. For example, heater power adjustments cannot exceed the rated rate of change, pulling speed adjustments cannot cause diameter control loop instability, and argon flow corrections cannot lead to furnace pressure exceeding limits. The system then converts executable measures into specific parameter adjustment instructions. These parameters can include corrections to heater power setpoints, crucible lift ratios, seed crystal lifting speeds, argon mass flow controller setpoints, furnace pressure control targets, and delay adjustment parameters for related components. If a single measure is insufficient to cover the root cause, the system can generate a set of candidate instructions for coordinated multi-measure action, assigning each set an execution priority, expected improvement target, and risk label. For example, in scenarios with thermal field fluctuations accompanied by atmosphere disturbances, candidate instructions could simultaneously include a combination of measures such as slightly adjusting heater power, fine-tuning argon flow, and reducing the slope of the pulling speed change to mitigate secondary disturbances caused by large adjustments to a single parameter. To improve consistency with field practices, the system can also modify the measure ranking using a human-machine collaboration mechanism. This involves adjusting the priority of candidate instructions based on engineers' historical acceptance rates, post-implementation compliance rates, and adaptation effects to similar operating conditions, but the output results remain based on the current root cause and safety constraints.

[0074] Based on the above analysis, this step explicitly maps the root causes of anomalies to historical experience and standard procedures. This ensures that anomaly handling no longer relies on subjective human judgment but generates candidate instructions with parameter-level executableness. Since these candidate instructions include adjustment targets, directions, magnitudes, and timing constraints, the control system can directly conduct subsequent simulation verification, avoiding the uncertain process of engineers having to decide "what to adjust, how much to adjust, and which to adjust first" after a traditional alarm. For systems like single-crystal furnaces with high thermal inertia and strong parameter coupling, considering the process stage, equipment capabilities, and safety boundaries during the measure generation phase helps reduce secondary anomalies caused by inappropriate handling and improves the stability and feasibility of subsequent execution plans.

[0075] S104. The candidate instructions are simulated and verified in the digital simulation model. If the simulation is successful, the candidate instructions are issued for execution. The digital simulation model is used to simulate the operating status of the single crystal furnace.

[0076] In this embodiment, the digital simulation model can be a digital twin model corresponding to the on-site single crystal furnace equipment, process formula, and control logic, used to simulate the impact of candidate instructions on the operating state of the single crystal furnace in a virtual environment. Simulation success means that the candidate instruction meets preset safety, feasibility, and effectiveness conditions during the simulation process. Safety conditions constrain variables such as temperature, pressure, actuator stroke, and load to not exceed allowable ranges. Feasibility conditions constrain the control instructions to be executed by the on-site equipment as required. Effectiveness conditions evaluate whether anomalies are suppressed and whether core process parameters tend to stabilize. Execution refers to sending the verified candidate instruction via the control interface to the target execution module in the HMI (Human-Machine Interface), PLC, DCS (Distributed Control System), or MES for direct execution by the automatic control unit, or execution after confirmation by the operator.

[0077] In practice, the system inputs each set of candidate instructions into the digital simulation model. The model uses the current real-time state of the single-crystal furnace as initial conditions and simulates the thermal field, flow field, pressure field, liquid level changes, and crystal growth state response within a predetermined time step within a future prediction window. The model can include heat conduction equations, fluid motion equations, boundary heat transfer conditions, actuator response models, and diameter control loop models, thereby reproducing the multi-physics linkage effects caused by heater power changes, pulling speed changes, gas flow rate changes, and crucible lifting actions. During the simulation, the system continuously evaluates multiple criteria, such as whether crystal diameter fluctuations converge, whether the temperature gradient returns to the allowable zone, whether the furnace pressure remains stable, whether the stress distribution avoids local overestimation, whether the actuator change rate is less than the equipment's allowable value, and whether the anomaly risk score is lower than a preset threshold. If a set of candidate instructions triggers any prohibitive conditions, such as predicting temperature exceeding limits, abnormal pressure, diameter instability amplification, or excessive structural stress, then that set of candidate instructions is judged as failing the simulation and automatically rolled back; the system can continue to select the next set of candidate instructions for simulation until a successful solution is obtained. For multiple candidate instructions that pass simultaneously, the system can sort them according to their overall deviation. The overall deviation can be composed of the weighted deviation of multiple key indicators relative to the target value; the smaller the deviation, the better the solution. After sorting, the system outputs the optimal candidate instruction and displays the simulation trend, key variable change curves, and risk assessment results to engineers through a human-machine interface. When the automatic execution conditions are met, the system can also omit manual confirmation and directly write the instruction into the controller setpoint area. The PLC then completes the control action according to the execution sequence and time parameters, while recording the execution time, execution results, and subsequent feedback data for iterative updates to the measure library and diagnostic model.

[0078] Based on the above analysis, this step, by verifying candidate instructions in a digital simulation model beforehand, ensures that the safety and effectiveness of handling anomalies in actual production are assessed before execution, thus avoiding the direct application of unverified adjustments to the single-crystal furnace. The Czochralski process for single-crystal silicon exhibits significant thermal inertia, hysteresis, and multivariate coupling characteristics. A minor parameter adjustment often takes several minutes or even longer to propagate to the crystal growth interface; therefore, relying solely on experience can easily lead to over-adjustment or directional deviation. This step utilizes a digital simulation model synchronized with the on-site conditions to predict future responses, identifying potential risks and automatically eliminating unsuitable solutions before execution, thereby improving the stability, continuity, and operational safety of the handling process.

[0079] Based on the above analysis, this application provides a control method for a single crystal furnace, including: collecting real-time operating data of the single crystal furnace and extracting key characteristic waveforms during abnormal periods; inputting the real-time operating data into a preset diagnostic model for anomaly identification; when an abnormal event is identified, performing root cause reasoning based on a knowledge graph to locate the root cause of the anomaly; retrieving historical measures and / or standard operating procedures that match the root cause of the anomaly from a measure library, and generating candidate instructions containing specific parameter adjustment instructions; performing simulation verification of the candidate instructions in a digital simulation model; and issuing the candidate instructions for execution when the simulation passes, wherein the digital simulation model is used to simulate the operating state of the single crystal furnace. In this embodiment, by connecting data acquisition, anomaly identification, root cause location, measure generation, simulation verification, and execution issuance into a closed-loop control process, rapid identification, accurate tracing, and stable handling of anomalies in the single crystal furnace are achieved. This transforms the traditional processing model, which relies on threshold alarms and manual experience analysis, into a collaborative decision-making model based on multi-source operational data, structured knowledge, and digital twin verification. This improves the lead time for anomaly identification, enhances the consistency of root cause analysis, reduces the blindness of response actions, and helps mitigate the adverse effects of crystal diameter instability, increased risk of wire breakage, and abnormal equipment linkage on product quality and continuous production. In one possible embodiment, the aforementioned diagnostic model, knowledge graph, measure library, and digital simulation model can be deployed on the same local industrial server, or distributed across edge computing nodes and manufacturing execution system servers, working collaboratively through industrial communication networks. It should be understood that the above examples are merely illustrative and not limiting.

[0080] The single crystal furnace control method provided in this application improves the timeliness of single crystal furnace anomaly identification and the accuracy of anomaly root cause location by collecting real-time operating data of the single crystal furnace and extracting key characteristic waveforms during abnormal periods, combining anomaly identification with a preset diagnostic model and root cause reasoning using a knowledge graph. By retrieving matching historical measures and / or standard operating procedures from the measure library based on the anomaly root cause, generating candidate instructions containing specific parameter adjustment instructions, and simulating and verifying the candidate instructions in a digital simulation model before issuing them for execution, the method can improve the pertinence and stability of anomaly handling decisions, thereby reducing reliance on human experience and improving the reliability of single crystal furnace operation control.

[0081] Figure 2 Flowchart of the control method for the single crystal furnace provided in this application Figure 2 ,like Figure 3 As shown, in this embodiment... Figure 2 Based on the embodiments, the control method of the single crystal furnace is described in detail, which includes:

[0082] S201. Clean, align, and extract features from real-time operating data, extracting at least one of the following during abnormal periods: oscillation decay waveform of crystal diameter, time-delay response waveform of liquid surface temperature, or periodic fluctuation waveform of pulling speed; extract morphological feature parameters of the waveforms; input the morphological feature parameters into a preset diagnostic model to identify abnormal events.

[0083] Among them, the morphological characteristic parameters include at least one of the following: oscillation frequency, attenuation coefficient, response lag, or fluctuation amplitude.

[0084] The real-time operational data includes time-series signals such as crystal diameter, liquid surface temperature, pulling speed, and furnace pressure, thermal power, and gas flow rate related to the operation of the single crystal furnace. Data cleaning is used to remove acquisition noise, missing segments, and sudden outliers. Alignment is used to eliminate curve misalignment caused by inconsistent sampling frequencies of different sensors and timestamp offsets. Feature extraction is used to extract abnormal periods from continuous time series and form waveforms to be identified. The oscillation decay waveform of the crystal diameter is used to characterize the decay and stabilization process of the diameter control quantity after disturbance. The time-delay response waveform of the liquid surface temperature is used to characterize the hysteresis characteristics of temperature change after temperature control adjustment. The periodic fluctuation waveform of the pulling speed is used to characterize the periodic oscillation phenomenon in the crystal pulling control loop.

[0085] In practical implementation, the original data can first be processed by median filtering, moving average, or wavelet denoising to reduce the impact of high-frequency interference on diagnosis. Then, based on a unified time reference, the multi-source data is resampled and interpolated to ensure that each curve corresponds on the same time axis. Subsequently, by combining the abnormal alarm points, threshold overruns, or deviations within the sliding window, the abnormal time period is located and the corresponding waveform is extracted. The morphological characteristic parameters of the waveform can be obtained through curve fitting, correlation analysis, or frequency domain analysis. For example, an exponential decay model can be fitted to the oscillation decay waveform to obtain the oscillation frequency and decay coefficient; the time difference between the input change and the output response can be calculated for the time-delay response waveform to obtain the response time delay; and the peak-to-valley difference or amplitude-frequency value can be calculated for the periodic fluctuation waveform to obtain the fluctuation amplitude.

[0086] After inputting morphological feature parameters into a pre-defined diagnostic model, the model comprehensively judges the stability, response speed, and control oscillation characteristics of the current waveform based on the trained anomaly pattern mapping relationship, thereby outputting the identification result of the anomaly event. This method can transform the implicit dynamic features in the original curve into quantifiable parameters, reduce the error of directly relying on manual image interpretation, and improve the ability to identify early anomalies and coupled anomalies. Since the pre-defined diagnostic model receives cleaned and aligned morphological feature parameters, its input data consistency is higher, thus improving the accuracy and real-time performance of anomaly judgment, reducing false negatives and false positives, and providing a stable data foundation for subsequent root cause tracing and handling decisions.

[0087] S202. Input the real-time operating data into the preset diagnostic model, wherein the preset diagnostic model includes a mechanism diagnostic sub-model and a data-driven diagnostic sub-model that run in parallel; wherein the mechanism diagnostic sub-model is constructed based on the thermal field mechanics equation or the fluid mechanics equation, and is used to determine from the perspective of physical mechanism whether the current operating state deviates from the theoretical expectation and output the first anomaly probability.

[0088] The data-driven diagnostic sub-model is a deep neural network model trained on historical data of line breakage accidents or quality anomalies. It is used to identify abnormal patterns from a data statistics perspective and output a second anomaly probability.

[0089] When at least one of the first and second abnormal probabilities exceeds a preset alarm threshold, an abnormal event is determined, and the abnormal type of the abnormal event is identified. Starting from the abnormal type, a multi-hop association query is performed in the knowledge graph using a reverse reasoning algorithm. The association is traced back layer by layer to calculate the conditional probability corresponding to each candidate root cause node, and the root cause node with the highest conditional probability is selected as the abnormal root cause.

[0090] The knowledge graph includes at least equipment nodes, process parameter nodes, raw material attribute nodes, and the relationships between them.

[0091] The system includes a pre-defined diagnostic model for parallel analysis of real-time operating data from the single crystal furnace; a mechanism-based diagnostic sub-model based on thermal field balance, pulling speed stability, gas flow coupling, and furnace pressure variation; and a data-driven diagnostic sub-model trained on historical anomaly samples to identify anomaly patterns from multi-parameter correlations. A knowledge graph characterizes the causal or correlational relationships between equipment status, process parameters, and raw material properties. Equipment nodes correspond to heaters, crucibles, flow guides, and temperature measuring components; process parameter nodes correspond to temperature, pulling speed, furnace pressure, and flow rate; and raw material property nodes correspond to raw material purity, charging status, and impurity levels.

[0092] In actual operation, real-time data, after sampling, cleaning, and normalization, is simultaneously input into two types of sub-models. The mechanism diagnosis sub-model outputs anomaly judgments based on preset mechanism thresholds and state equations, while the data-driven diagnosis sub-model outputs anomaly signals based on classification probabilities or anomaly scores. When either result meets the confidence requirement, an anomaly event is output and mapped to types such as line breakage risk, diameter fluctuation, temperature anomaly, or oxygen content anomaly. Subsequently, using the anomaly type as the retrieval entry point, the system searches backward in the knowledge graph for process parameters, equipment status, and raw material characteristics associated with that type. Combining edge weights, historical co-occurrence counts, and Bayesian conditional probabilities, the system traces the root cause chain layer by layer, ultimately selecting the node with the highest probability as the root cause of the anomaly to provide a basis for subsequent measure matching and parameter adjustment.

[0093] This approach improves the robustness and timeliness of anomaly identification by running mechanistic diagnosis and data-driven diagnosis in parallel; through multi-hop reverse reasoning of knowledge graphs, it can establish traceable associations between anomaly types and potential causes, reducing reliance on human experience; by screening root cause nodes through conditional probability, it can reduce the probability of misjudgment and improve the consistency of localization, thereby enhancing the accuracy of single crystal furnace anomaly diagnosis and the stability of handling.

[0094] S203. Input the candidate instruction into the digital simulation model and simulate the parameter adjustment action corresponding to the candidate instruction. The digital simulation model is used to simulate the predicted operating state of the single crystal furnace under the candidate instruction. Obtain at least one predicted operating index output by the digital simulation model after the simulation is executed, and determine whether the predicted operating index meets the preset safety and stability conditions. If the safety and stability conditions are met, the simulation is deemed to have passed, and the candidate instruction is allowed to be issued for execution. If the safety and stability conditions are not met, the simulation is deemed to have failed, the candidate instruction is abandoned, and alternative instructions are retrieved from the measure library for re-simulation verification.

[0095] The digital simulation model can be constructed as a digital twin model corresponding to the single crystal furnace control object. It pre-sets state variables such as the furnace thermal field, crucible liquid surface, crystal growth interface, gas flow field, and control response. Model parameters can be jointly calibrated based on historical production data, equipment calibration data, and process mechanism parameters to ensure the model reflects the predicted operating state under candidate commands. After a candidate command is input, the simulation module adjusts and updates the heating power, pulling speed, argon flow rate, furnace pressure, or crucible lifting position according to the parameters corresponding to the command. It calculates the state evolution results within a preset time window and outputs predicted operating indicators such as crystal diameter fluctuation amplitude, temperature gradient, liquid surface stability, thermal field uniformity, or control quantity change stability. Pre-set safety and stability conditions can be set such as a single indicator not exceeding a threshold, multiple indicators falling within the allowable range simultaneously, or the comprehensive evaluation value satisfying constraints, to ensure that the candidate command will not cause significant thermal field instability or control overshoot after actual issuance.

[0096] After acquiring the simulation output, the system performs a consistency check on the predicted operating indicators and compares the check results with the safety and stability conditions. When the comparison result shows that the requirements are met, the control unit marks the candidate instruction as an executable instruction and sends it to the single crystal furnace actuator so that the corresponding parameter adjustment can be completed on the real equipment. When the comparison result shows that the requirements are not met, the system marks the current candidate instruction as a failed instruction and stops issuing it. At the same time, it retrieves alternative instructions from the measure library based on the root cause of the anomaly or process constraints, and drives the digital simulation model again for verification, thereby forming a closed loop for candidate instruction screening. In practical applications, the measure library can also adopt other organizational forms, which are not limited in this embodiment.

[0097] This method involves simulating and screening candidate commands before issuing them. This allows the control system to evaluate the operational consequences of parameter adjustments in a virtual environment before deciding whether to execute them on-site. This reduces the risks of crystal diameter fluctuations, thermal imbalances, and equipment impacts caused by unreasonable adjustments, and improves the stability and consistency of anomaly handling. Since alternative commands can be automatically switched for further verification when the simulation fails, repeated manual trial and error are reduced, improving efficiency and enhancing the safety and controllability of continuous single-crystal furnace production.

[0098] For example, multiple candidate instructions are retrieved from the measure library, and the predicted operating indicators of each candidate instruction after execution in the digital simulation model are estimated. Based on the comprehensive deviation between each predicted operating indicator and its corresponding target operating indicator, the candidate instructions are sorted, and the candidate instruction with the smallest comprehensive deviation is selected as the new candidate instruction. The new candidate instruction is re-input into the digital simulation model for simulation verification until a candidate instruction that passes the simulation is found or all candidate instructions have been retrieved.

[0099] If the safety and stability conditions are not met, the simulation is deemed unsuccessful, the candidate instruction is discarded, and alternative instructions are retrieved from the measure library for re-simulation verification. The method also includes: retrieving multiple alternative instructions from the measure library, estimating the predicted operating indicators of each alternative instruction after execution in the digital simulation model; ranking the alternative instructions according to the comprehensive deviation between each predicted operating indicator and its corresponding target operating indicator, and selecting the alternative instruction with the smallest comprehensive deviation as the new candidate instruction; re-inputting the new candidate instruction into the digital simulation model for simulation verification, until a candidate instruction that passes the simulation is found or all alternative instructions have been retrieved.

[0100] The measures library stores historical handling measures, standard operating procedures, and parameter correction instructions corresponding to the root causes of single crystal furnace anomalies. Alternative instructions may include heating power correction values, pulling speed correction values, gas flow compensation values, or combinations thereof. The digital simulation model characterizes the temperature field, liquid level state, crystal pulling stability, and crystal quality evolution of the single crystal furnace under different instructions. It can be calibrated based on historical operating data, mechanistic parameters, and equipment boundary conditions to improve the consistency between predicted results and actual operating conditions. In this embodiment, the target operating index can be taken from process baseline values, standard operating windows, or control thresholds set by expert experience under stable production conditions. In practical applications, the target operating index can also be selected from other sources; this embodiment does not limit this.

[0101] In practical implementation, the system retrieves multiple candidate instructions from the measure library with varying degrees of matching to the root causes of anomalies. Each candidate instruction is mapped to a corresponding parameter adjustment combination and input into the digital simulation model for preliminary calculation, yielding the corresponding predicted operating indicators. These predicted operating indicators may include at least temperature gradient, crucible liquid level fluctuation amplitude, pulling speed stability, furnace pressure deviation, crystal diameter fluctuation, and alarm risk value. Subsequently, the system compares each predicted operating indicator with its corresponding target operating indicator and weights the differences to obtain the overall deviation degree. The weighting coefficients can be pre-configured based on process safety, control priority, and anomaly type to enhance sensitivity to deviations in key indicators. The system sorts the candidate instructions from smallest to largest overall deviation degree, thus prioritizing the candidate instruction closest to the target operating state as the new candidate instruction.

[0102] After a new candidate instruction is selected, the system inputs it into the digital simulation model again for simulation verification and determines whether the corresponding predicted operating indicators meet the safety and stability conditions. If the simulation passes, the executable candidate instruction is output and proceeds to the subsequent issuance stage; if it still fails, the system continues to search for unverified alternative instructions from the measure library and repeats the estimation, sorting, and verification process until a passable candidate instruction is obtained or all alternative instructions have been retrieved. Through this iterative screening method, the system can prioritize and lock in the adjustment scheme with the smallest deviation within a limited range of measures.

[0103] By adopting the above method, when a candidate instruction fails, the system can automatically switch to a progressive screening process using multiple alternative instructions, avoiding simulation failures or misjudgments caused by relying on a single measure. Because a comprehensive deviation evaluation mechanism between predicted and target operating indicators is introduced, the accuracy and hit rate of instruction screening can be improved, the probability of incompatible instructions being executed on-site can be reduced, and the stability and continuity of single-crystal furnace anomaly handling can be enhanced.

[0104] For example, a diagnostic report and simulation verification results are pushed to the engineer through a human-machine interface, and a manual confirmation timer is started. If the engineer's confirmation signal is received within a preset time, the candidate instruction is sent to the programmable logic controller or manufacturing execution system through the industrial communication protocol in response to the confirmation signal to modify the process setting value. If no confirmation signal is received within the preset time, the candidate instruction is automatically sent to the programmable logic controller or manufacturing execution system through the industrial communication protocol.

[0105] The human-machine interface (HMI) can be a host computer interface for a single-crystal furnace, an industrial touchscreen, or a remote monitoring terminal. The diagnostic report presents the root cause of the anomaly, candidate instruction content, simulation verification results, and an explanation of the impact on process setpoints. A manual confirmation timer records the engineer's response time and can be triggered immediately by the control server after a candidate instruction is generated. Confirmation signals can be generated after the engineer clicks a confirmation button on the interface, enters an authorization password, or completes an electronic signature, ensuring the traceability of issued operations.

[0106] In practice, industrial communication protocols can employ OPC UA (Open Platform Communications Unified Architecture), Modbus, or PROFINET (Process Field Network) to convert candidate instructions into control messages recognizable by the PLC or MES, and further map them into modification commands for process settings such as heating power, pulling speed, gas flow rate, furnace pressure, or crucible lifting speed. Upon receiving the instruction, the PLC will write the parameters according to the current interlock status and access control logic. Upon receiving the instruction, the MES will synchronize the candidate instruction to the work order or recipe management module to maintain consistency in process records.

[0107] This implementation introduces a parallel issuance mechanism of manual confirmation and automatic execution upon timeout after simulation verification. This allows engineers to perform manual review in high-risk scenarios while maintaining the continuity of the control loop even if confirmation is not received within a preset time. This mechanism avoids delays caused by manual waiting during anomaly handling and reduces the risk of missing the optimal adjustment opportunity due to untimely execution of candidate instructions, thereby improving the real-time performance, stability, and traceability of single-crystal furnace process adjustments.

[0108] For example, actual operational data after execution is collected, and an operation log for this anomaly handling is recorded. The operation log includes: anomaly type, the identified root cause of the anomaly, the specific parameters of the candidate instructions, whether the engineer modified the candidate instructions, the content of the modified instructions, and the evaluation of the control effect after execution. The operation log is automatically stored in the knowledge base. When the preset trigger conditions are met, the data-driven diagnostic sub-model is incrementally trained using the newly added data in the knowledge base, the entity relationships and attributes of the knowledge graph are updated, and the historical measures in the measure library and their application effect scores are dynamically adjusted.

[0109] The actual operating data can include temperature curves, pulling speed curves, gas flow rate curves, crystal diameter changes, and equipment status feedback after the execution of candidate commands. Control effectiveness evaluation can be determined by stability, response speed, deviation convergence degree, and product quality indicators. Operation logs are automatically generated by the control terminal after the candidate command is executed and stored in association with the timestamp of the corresponding abnormal event, furnace number, process batch, and command version number for subsequent traceability. The data-driven diagnostic sub-model can use online learning or small-batch incremental updates, using the abnormality type labels, root cause labels, and effectiveness evaluations from the newly added logs as training samples to keep the model adaptable to new operating conditions. Entity relationships and attribute updates in the knowledge graph include correcting the association weights between equipment nodes, process parameter nodes, raw material attribute nodes, and measure nodes, and supplementing the mapping relationship between abnormality types and root causes based on newly added logs. Historical measures and their application effect scores in the measure library can be re-evaluated based on the control effectiveness evaluation after execution, thereby improving the accuracy of subsequent measure retrieval and recommendation.

[0110] This implementation method collects, records, and stores the results of anomaly handling, creating a closed-loop data system that includes the execution effects of candidate instructions, traces of manual modifications, and root cause information. Furthermore, by synchronously updating the diagnostic sub-model, knowledge graph, and measure library in conjunction with triggering conditions, subsequent anomaly identification, root cause localization, and measure recommendation can be continuously iterated and optimized, reducing reliance on single-instance experience-based judgments.

[0111] By adopting this implementation method, effective experience in the abnormal handling process can be automatically transformed into reusable knowledge, which can improve the ability of the data-driven diagnostic sub-model to identify new abnormal patterns and enhance the consistency between the knowledge graph reasoning results and the actual situation on site. At the same time, it can keep the historical measure scores in the measure library synchronized with the actual control effect, thereby improving the accuracy, stability and response efficiency of subsequent abnormal handling.

[0112] The single crystal furnace control method provided in this application improves the timeliness of single crystal furnace anomaly identification and the accuracy of anomaly root cause location by collecting real-time operating data of the single crystal furnace and extracting key characteristic waveforms during abnormal periods, combining anomaly identification with a preset diagnostic model and root cause reasoning using a knowledge graph. By retrieving matching historical measures and / or standard operating procedures from the measure library based on the anomaly root cause, generating candidate instructions containing specific parameter adjustment instructions, and simulating and verifying the candidate instructions in a digital simulation model before issuing them for execution, the method can improve the pertinence and stability of anomaly handling decisions, thereby reducing reliance on human experience and improving the reliability of single crystal furnace operation control.

[0113] Figure 4 This is a schematic diagram of the control device for the single crystal furnace provided in this application, as shown below. Figure 4As shown, the control device 40 for the single crystal furnace provided in this embodiment includes:

[0114] The acquisition module 401 is used to acquire real-time operating data of the single crystal furnace and extract key characteristic waveforms during abnormal periods of the single crystal furnace.

[0115] The identification module 402 is used to input real-time running data into a preset diagnostic model for anomaly identification. When an abnormal event is identified, the root cause is located by using a knowledge graph for root cause reasoning.

[0116] The matching module 403 is used to retrieve matching historical measures and / or standard operating procedures from the measures library based on the root cause of the anomaly, and generate candidate instructions containing specific parameter adjustment instructions;

[0117] The execution module 404 is used to perform simulation verification of candidate instructions in the digital simulation model. If the simulation is successful, the candidate instructions are issued for execution. The digital simulation model is used to simulate the operating state of the single crystal furnace.

[0118] In one possible implementation, the identification module 402 is used to input real-time running data into a preset diagnostic model, wherein the preset diagnostic model includes a mechanism diagnostic sub-model and a data-driven diagnostic sub-model that run in parallel.

[0119] Among them, the mechanism diagnosis sub-model is constructed based on the thermal field mechanics equation or the fluid mechanics equation, and is used to determine whether the current operating state deviates from the theoretical expectation from the perspective of physical mechanism, and output the first anomaly probability;

[0120] The data-driven diagnostic sub-model is a deep neural network model trained on historical data of line breakage accidents or quality anomalies. It is used to identify abnormal patterns from a data statistics perspective and output a second anomaly probability.

[0121] When at least one of the first abnormal probability and the second abnormal probability exceeds the preset alarm threshold, it is determined that an abnormal event exists and the abnormal type of the abnormal event is determined.

[0122] Starting with the anomaly type, a reverse reasoning algorithm is used to perform multi-hop association queries in the knowledge graph. The knowledge graph contains at least equipment nodes, process parameter nodes, raw material attribute nodes, and the relationships between the three.

[0123] By tracing back layer by layer along the association relationship, the conditional probability corresponding to each candidate root cause node is calculated, and the root cause node with the highest conditional probability is selected as the abnormal root cause.

[0124] In one possible implementation, the execution module 404 is used to input candidate instructions into the digital simulation model and simulate the parameter adjustment action corresponding to the candidate instructions. The digital simulation model is used to simulate the predicted operating state of the single crystal furnace under the candidate instructions.

[0125] Obtain at least one predicted operating index output by the digital simulation model after the simulation is executed, and determine whether the predicted operating index meets the preset safety and stability conditions.

[0126] If the safety and stability conditions are met, the simulation is deemed successful, and the candidate instructions can be issued for execution.

[0127] If the safety and stability conditions are not met, the simulation is deemed unsuccessful, the candidate instructions are abandoned, and alternative instructions are retrieved from the measure library for re-simulation verification.

[0128] In one possible implementation, the execution module 404 is used to retrieve multiple alternative instructions from the measure library and estimate the various predicted operating indicators after each alternative instruction is executed in the digital simulation model.

[0129] Based on the overall deviation between each predicted operating indicator and its corresponding target operating indicator, the candidate instructions are ranked, and the candidate instruction with the smallest overall deviation is selected as the new candidate instruction.

[0130] The new candidate instructions are re-input into the digital simulation model for simulation verification until a candidate instruction that passes the simulation is found or all alternative instructions have been retrieved.

[0131] In one possible implementation, the execution module 404 is used to push diagnostic reports and simulation verification results to the engineer through a human-machine interface and start a manual confirmation timer;

[0132] If an acknowledgment signal from the engineer is received within a preset time, the candidate instruction is sent to the programmable logic controller or manufacturing execution system via the industrial communication protocol in response to the acknowledgment signal to modify the process settings.

[0133] If no confirmation signal is received within the preset time, the candidate instruction will be automatically sent to the programmable logic controller or manufacturing execution system via the industrial communication protocol.

[0134] In one possible implementation, the acquisition module 401 is used to clean, align and extract features from the real-time running data, and extract at least one of the following during abnormal periods: the oscillation decay waveform of the crystal diameter, the time-delay response waveform of the liquid surface temperature or the periodic fluctuation waveform of the pulling speed.

[0135] Extract the waveform's morphological feature parameters, which include at least one of the following: oscillation frequency, attenuation coefficient, response delay, or fluctuation amplitude.

[0136] Input morphological feature parameters into a preset diagnostic model to identify abnormal events.

[0137] In one possible implementation, the execution module 404 is used to collect the actual running data after execution and record the operation log of this exception handling. The operation log includes: exception type, the root cause of the exception located, the specific parameters of the candidate instruction, whether the engineer has modified the candidate instruction, the content of the modified instruction, and the evaluation of the control effect after execution.

[0138] Automatically save operation logs to the knowledge base;

[0139] When the preset triggering conditions are met, the data-driven diagnostic sub-model is incrementally trained using the newly added data in the knowledge base, the entity relationships and attributes of the knowledge graph are updated, and the historical measures in the measures library and their application effect scores are dynamically adjusted.

[0140] The control device for the single crystal furnace provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0141] Figure 4 This is a schematic diagram of the control equipment for the single crystal furnace provided in this application. Figure 4 As shown, the electronic device 50 provided in this embodiment includes at least one processor 501 and a memory 502. Optionally, the device 50 further includes a communication component 503. The processor 501, memory 502, and communication component 503 are connected via a bus 504.

[0142] In a specific implementation, at least one processor 501 executes computer execution instructions stored in memory 502, causing at least one processor 501 to perform the above-described method.

[0143] The specific implementation process of processor 501 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0144] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0145] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0146] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0147] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0148] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0149] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0150] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0151] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0152] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0153] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0154] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0155] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0156] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A method for controlling a single crystal furnace, characterized in that, include: Collect real-time operating data of the single crystal furnace and extract key characteristic waveforms during abnormal periods of the single crystal furnace; The real-time operating data is input into a preset diagnostic model for anomaly identification. When an abnormal event is identified, the root cause is located by using a knowledge graph for root cause reasoning. Based on the root cause of the anomaly, retrieve matching historical measures and / or standard operating procedures from the measures library to generate candidate instructions containing specific parameter adjustment instructions; The candidate instructions are simulated and verified in a digital simulation model. If the simulation passes, the candidate instructions are issued for execution. The digital simulation model is used to simulate the operating state of the single crystal furnace.

2. The method according to claim 1, characterized in that, The real-time operational data is input into a preset diagnostic model for anomaly identification. When an abnormal event is identified, root cause reasoning is performed using a knowledge graph to locate the root cause of the anomaly, including: The real-time running data is input into the preset diagnostic model, wherein the preset diagnostic model includes a mechanism diagnostic sub-model and a data-driven diagnostic sub-model that run in parallel. The mechanism diagnosis sub-model is constructed based on the thermal field mechanics equation or the fluid mechanics equation, and is used to determine whether the current operating state deviates from the theoretical expectation from the perspective of physical mechanism, and output the first anomaly probability. The data-driven diagnostic sub-model is a deep neural network model trained based on historical data of wire breakage accidents or quality anomalies. It is used to identify abnormal patterns from a data statistics perspective and output a second anomaly probability. When at least one of the first anomaly probability and the second anomaly probability exceeds a preset alarm threshold, an anomaly event is determined to exist, and the anomaly type of the anomaly event is determined. Starting with the aforementioned anomaly type, a reverse reasoning algorithm is used to perform multi-hop association queries in the knowledge graph, wherein the knowledge graph includes at least equipment nodes, process parameter nodes, raw material attribute nodes, and the relationships between the three. By tracing back layer by layer along the aforementioned relationship, the conditional probability corresponding to each candidate root cause node is calculated, and the root cause node with the highest conditional probability is selected as the abnormal root cause.

3. The method according to claim 1, characterized in that, The candidate instructions are simulated and verified in a digital simulation model, including: The candidate instruction is input into the digital simulation model to simulate the parameter adjustment action corresponding to the candidate instruction. The digital simulation model is used to simulate the predicted operating state of the single crystal furnace under the candidate instruction. Obtain at least one predicted operating index output by the digital simulation model after the simulation is executed, and determine whether the predicted operating index meets the preset safety and stability conditions. If the aforementioned safety and stability conditions are met, the simulation is deemed successful, and the candidate instructions can be issued for execution. If the aforementioned safety and stability conditions are not met, the simulation is deemed unsuccessful, the candidate instruction is discarded, and alternative instructions are retrieved from the measure library for re-simulation verification.

4. The method according to claim 3, characterized in that, If the aforementioned safety and stability conditions are not met, the simulation is deemed unsuccessful, the candidate instruction is discarded, and alternative instructions are retrieved from the measure library for re-simulation verification. The method further includes: Multiple alternative instructions are retrieved from the measure library, and the predicted operating indicators after each alternative instruction is executed in the digital simulation model are estimated. Based on the comprehensive deviation between each predicted operating indicator and its corresponding target operating indicator, the candidate instructions are sorted, and the candidate instruction with the smallest comprehensive deviation is selected as the new candidate instruction. The new candidate instructions are re-input into the digital simulation model for simulation verification until a candidate instruction that passes the simulation is found or all alternative instructions have been retrieved.

5. The method according to any one of claims 1 to 4, characterized in that, The candidate instruction is then issued and executed. The method further includes: The diagnostic report and simulation verification results are pushed to the engineer through the human-machine interface, and a manual confirmation timer is started; If an engineer's confirmation signal is received within a preset time, the candidate instruction is sent to the programmable logic controller or manufacturing execution system via the industrial communication protocol in response to the confirmation signal to modify the process setting value. If the confirmation signal is not received within a preset time, the candidate instruction will be automatically sent to the programmable logic controller or manufacturing execution system via the industrial communication protocol.

6. The method according to any one of claims 1 to 4, characterized in that, After collecting the real-time operating data of the single crystal furnace, the method further includes: The real-time operating data is cleaned, aligned, and feature extracted to extract at least one of the following during abnormal periods: the oscillation decay waveform of the crystal diameter, the time-delay response waveform of the liquid surface temperature, or the periodic fluctuation waveform of the pulling speed. Extract the morphological feature parameters of the waveform, wherein the morphological feature parameters include at least one of the following: oscillation frequency, attenuation coefficient, response delay, or fluctuation amplitude; The morphological feature parameters are input into the preset diagnostic model to identify abnormal events.

7. The method according to any one of claims 1 to 4, characterized in that, After the candidate instruction is issued and executed, the method further includes: Collect actual running data after execution and record the operation log of this exception handling. The operation log includes: exception type, root cause of exception located, specific parameters of the candidate instruction, whether the engineer modified the candidate instruction, content of the modified instruction, and evaluation of the control effect after execution. The operation log will be automatically stored in the knowledge base; When the preset triggering conditions are met, the data-driven diagnostic sub-model in the preset diagnostic model is incrementally trained using the newly added data in the knowledge base, the entity relationships and attributes of the knowledge graph are updated, and the historical measures and their application effect scores in the measures library are dynamically adjusted.

8. A control device for a single crystal furnace, characterized in that, include: The acquisition module is used to acquire real-time operating data of the single crystal furnace and extract key characteristic waveforms during abnormal periods of the single crystal furnace. The identification module is used to input the real-time running data into a preset diagnostic model for anomaly identification. When an abnormal event is identified, the root cause is located by using a knowledge graph for root cause reasoning. The matching module is used to retrieve matching historical measures and / or standard operating procedures from the measures library based on the root cause of the anomaly, and generate candidate instructions containing specific parameter adjustment instructions. The execution module is used to perform simulation verification of the candidate instructions in the digital simulation model. If the simulation passes, the candidate instructions are issued for execution. The digital simulation model is used to simulate the operating state of the single crystal furnace.

9. A control device for a single crystal furnace, characterized in that, include: Memory, processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1-7.

11. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method described in any one of claims 1-7.

Images