Reduction stove starting time prediction method, device, program and storage medium

Abstract

The invention discloses a restoration stove starting time prediction method and device, a program and a storage medium. The method comprises the following steps: acquiring process parameter time sequence data and a historical database of a target reduction furnace; formatting the process parameter time sequence data, and outputting a time sequence matrix and a current feature vector; according to the current feature vector and a preset pre-cleaning judgment strategy, whether pre-cleaning of the target reduction furnace is completed or not is determined; if so, inputting the time sequence matrix and a preset parameter set into a duration prediction model, and outputting a first prediction duration; determining a current process step of the target reduction furnace according to the current feature vector and a preset starting judgment strategy; according to the current process step, matching a second prediction duration required by the target reduction furnace from the current process step to the starting state from the historical database; and determining the final starting time based on the historical database, the first predicted duration and the second predicted duration. According to the invention, the accuracy of predicting and restoring the starting time of the stove can be improved.

Description

Technical Field

[0001] This application relates to the field of polysilicon production technology, specifically to a method, equipment, program, and storage medium for predicting the start-up time of a reduction furnace. Background Technology

[0002] The modified Siemens process is currently the mainstream technology for industrial polysilicon production. As a high-energy-consuming industry, electricity costs are a key factor restricting cost reduction and efficiency improvement. To comply with national time-of-use electricity pricing policies and reduce production costs, companies typically need to finely schedule high-energy-consuming reduction production processes, striving to avoid peak electricity price periods and make better use of off-peak hours. To achieve this goal, the scheduling system needs to accurately grasp the operating status and time nodes of each reduction furnace for coordinated operation. Currently, the technology for monitoring and analyzing reduction furnaces in operation (online status) through real-time data acquisition is relatively mature and can effectively support production decisions.

[0003] However, when the reduction furnace is offline—that is, in the preparation stage from shutdown cleaning, inspection, airtightness testing, and gas replacement to meet start-up conditions—there are still significant technical shortcomings in predicting the time required for startup. Under the current production model, determining when the reduction furnace will complete the preparation process and enter the "start-up ready" state relies primarily on the manual experience of the workshop operators. Operators need to manually observe parameters such as pressure and flow on instruments, combine this with their subjective experience to estimate the remaining time, and then manually enter the predicted results into the system.

[0004] This traditional manual forecasting method has many drawbacks, severely impacting the accuracy of start-up time predictions. First, the preparation process of the reduction furnace involves a series of complex physical and chemical processes such as water supply, gas supply, and purging, with parameter changes exhibiting non-linear characteristics, making precise quantification difficult solely based on human experience. Second, differences in experience levels, judgment standards, and sense of responsibility among different shifts and operators lead to highly random and subjective prediction results, lacking a unified standard. Third, manual reporting inherently has a time lag, failing to reflect sudden changes in on-site conditions in real time. These factors collectively result in low accuracy and significant data deviations in the current technology for predicting reduction furnace start-up times, making the future status information obtained by the production scheduling system unreliable. Consequently, production scheduling plans become disconnected from actual execution, hindering the seamless integration of production processes and optimal control of electricity costs. Summary of the Invention

[0005] The purpose of this application is to provide a method, device, program, and storage medium for predicting the start-up time of a reduction furnace, so as to solve the problem of low accuracy in predicting the start-up time of a reduction furnace in the prior art.

[0006] To achieve the above objectives, the first aspect of this application provides a method for predicting the start-up time of a reduction furnace, the method comprising: Obtain the timing data and historical database of the process parameters of the target reduction furnace; The process parameter time series data is formatted and outputs a time series matrix and the current feature vector. Determine whether the target reduction furnace has completed pre-cleaning based on the current feature vector and the preset pre-cleaning judgment strategy; If the pre-cleaning has been completed, the time series matrix and the preset parameter set are input into the pre-built duration prediction model, and the first predicted duration is output. The current process step of the target reduction furnace is determined based on the current feature vector and the preset start-up judgment strategy; the preset start-up judgment strategy stores the mapping relationship between the feature vector and the process step; Based on the current process steps, a second predicted time is determined from the historical database to get the target reduction furnace from the current process steps to the start-up state. The final start time is determined based on historical databases, the first prediction duration, and the second prediction duration.

[0007] In this embodiment, the current feature vector includes the following vector elements: the current ratio of the furnace water regulating valve, the current temperature and current flow rate of the furnace water return, the current opening degree of the hydrogen regulating valve, the current opening degree of the trichlorosilane regulating valve, the current opening degree of the nitrogen regulating valve, and the current pressure inside the target reduction furnace; the preset pre-cleaning judgment strategy includes: a preset error threshold set, the target ratio of the furnace water regulating valve, the target temperature and target flow rate of the furnace water return, the target opening degree of the hydrogen regulating valve, the target opening degree of the trichlorosilane regulating valve, the target opening degree of the nitrogen regulating valve, and the target pressure inside the target reduction furnace; the preset error threshold set includes a first error threshold, a second error threshold, and a third error threshold.

[0008] In this embodiment of the application, the step of determining whether the target reduction furnace has completed pre-cleaning based on the current feature vector and the preset pre-cleaning judgment strategy includes: determining the status of the water supply and exhaust process of the target reduction furnace based on the current feature vector and the preset pre-cleaning judgment strategy; determining the status of the nitrogen replacement process of the target reduction furnace based on the current feature vector and the preset pre-cleaning judgment strategy; and determining that the target reduction furnace has completed pre-cleaning when both the water supply and exhaust process and the nitrogen replacement process are in a completed state.

[0009] In this embodiment of the application, the step of determining the state of the water supply and exhaust process of the target reduction furnace based on the current feature vector and the preset pre-cleaning judgment strategy includes: determining whether the absolute value of the difference between the current ratio and the target ratio of the furnace water regulating valve is less than a first error threshold; if it is less than the first error threshold, determining whether the absolute value of the difference between the current temperature and the target temperature of the furnace water return is less than or equal to the first error threshold; determining whether the absolute value of the difference between the current flow rate and the target flow rate of the furnace water return is less than or equal to the first error threshold; if all the above conditions are met, the water supply and exhaust process of the target reduction furnace is determined to be in a completed state.

[0010] In this embodiment of the application, the step of determining the state of the nitrogen replacement process of the target reduction furnace based on the current feature vector and the preset pre-cleaning judgment strategy includes: sequentially judging whether the following conditions are met, wherein the step of judging whether the latter condition is met is performed if the former condition is met: determining whether the absolute value of the difference between the current opening degree of the hydrogen regulating valve and the target opening degree of the hydrogen regulating valve is less than or equal to a second error threshold; determining whether the absolute value of the difference between the current opening degree of the trichlorosilane regulating valve and the target opening degree of the trichlorosilane regulating valve is less than or equal to a second error threshold; determining whether the absolute value of the difference between the current opening degree of the nitrogen regulating valve and the target opening degree of the nitrogen regulating valve is less than or equal to a second error threshold; determining whether the absolute value of the difference between the current pressure and the target pressure in the target reduction furnace is less than or equal to a third error threshold; determining whether the latest current opening degree of the nitrogen regulating valve is less than or equal to the second error threshold; if all the above conditions are met, the nitrogen replacement process is determined to be in a completed state.

[0011] In this embodiment, the duration prediction model includes a long short-term memory model, a dropout layer, and a fully connected layer. The long short-term memory model includes a forget gate and an input gate. The step of inputting a time series matrix and a preset parameter set into the pre-constructed duration prediction model and outputting a first predicted duration includes: determining the cell state data at the current time based on the time series matrix, the hidden state data and cell state data output by the duration prediction model at the previous time step, and the preset parameter set; determining the context feature vector based on the time series matrix, the hidden state data output by the duration prediction model at the previous time step, the cell state data at the current time step, and the preset parameter set; inputting the context feature vector into the dropout layer to reduce overfitting and outputting the processed feature data; inputting the feature data into the fully connected layer for calculation, and activating the calculation result through the ReLU function to output the first predicted duration.

[0012] In this embodiment, the Long Short-Term Memory (LSTM) model includes a forget gate and an input gate; the preset parameter set includes: preset trainable weight parameters, a preset forget gate bias vector, a preset input gate bias vector, a preset input gate weight matrix, a preset candidate state weight matrix, and a preset candidate state bias vector; the step of determining the cell state data at the current time based on the time series matrix, the hidden state data and cell state data output by the duration prediction model at the previous time step, and the preset parameter set includes: combining the time series matrix, the hidden state data output by the duration prediction model at the previous time step, the preset trainable weight parameters, and the preset forget gate bias vector. Input the input vector to the forget gate and output the forget gate vector data; input the time series matrix, the hidden state data output by the duration prediction model at the previous time step, the preset input gate bias vector, and the preset input gate weight matrix to the input gate and output the input gate vector data; determine the candidate state data based on the time series matrix, the hidden state data output by the duration prediction model at the previous time step, the preset candidate state weight matrix, and the preset candidate state bias vector; determine the cell state data at the current time step based on the candidate state data, the cell state data output by the duration prediction model at the previous time step, the forget gate vector data, and the input gate vector data.

[0013] In this embodiment, the duration prediction model further includes an attention layer; the long short-term memory model further includes an output gate; the preset parameter set includes: preset output gate weights, preset output gate bias parameters, preset attention weight parameter set, and preset attention bias parameters; the step of determining the context feature vector based on the time series matrix, the hidden state data output by the duration prediction model at the previous time step, the cell state data at the current time step, and the preset parameter set includes: inputting the preset output gate weights, preset output gate bias parameters, the time series matrix, and the hidden state data output by the duration prediction model at the previous time step into the output gate to obtain output gate vector data; determining the hidden state data at the current time step based on the output gate vector data and the cell state data at the current time step; inputting the hidden state data at the current time step, the preset attention weight parameter set, and the preset attention bias parameters into the attention layer to output the attention score at the current time step; determining the normalized weight corresponding to the attention score at the current time step based on the sum of the attention scores at the current time step and all historical time steps; and determining the context feature vector based on the normalized weight and the hidden state data at the current time step.

[0014] In this embodiment, the historical database includes historical production batch data; matching the target reduction furnace from the current process step to the start-up state from the historical database according to the current process step includes determining the second prediction time according to the following formula:

[0015] in, This is the second prediction duration; This represents the total number of times the current process step appears in historical production batch data. This refers to the actual time interval between the kth occurrence of the current process step in the historical production batch data and the start-up status of the tool. This represents the current process step.

[0016] In this embodiment of the application, the step of determining the final start time based on the historical database, the first predicted duration, and the second predicted duration includes: determining a first difference between the first predicted duration and the corresponding actual duration in the historical production batch data; determining a second difference between the second predicted duration and the corresponding actual duration in the historical production batch data; determining a first weight of the first predicted duration and a second weight of the second predicted duration based on the first difference and the second difference; and determining the final start time based on the first weight, the second weight, the first predicted duration, and the second predicted duration.

[0017] In this embodiment of the application, determining the first difference between the first predicted duration and the corresponding actual duration in the historical production batch data includes: determining the first difference according to the following formula:

[0018] Determining the second difference between the second predicted duration and the corresponding actual duration in historical production batch data includes: determining the second difference according to the following formula:

[0019] in, This is the first difference; The second difference; n is the total number of batches in the historical production batch data; The first in the historical production batch data Within each batch, the predicted duration generated by the duration prediction model; The first in the historical production batch data Within each batch, the predicted duration is obtained by matching with the historical database; The first in the historical production batch data Within each batch, the actual time interval between the current process step and the start-up state.

[0020] In this embodiment of the application, determining the first weight and the second weight of the first prediction duration based on the first difference and the second difference includes: determining the first weight and the second weight according to the following formula:

[0021]

[0022]

[0023] in, It is the first weight; As the second weight; This is the first difference; This is the second difference.

[0024] In this embodiment of the application, determining the final start time based on the first weight, the second weight, the first prediction duration, and the second prediction duration includes: determining the final start time according to the following formula:

[0025] in, For the final start time; It is the first weight; The first prediction duration; As the second weight; This is the second prediction duration.

[0026] A second aspect of this application provides a computer device, comprising: The memory is configured to store instructions; and The processor is configured to retrieve instructions from memory and to implement the methods described above when executing instructions.

[0027] A third aspect of this application provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0028] A fourth aspect of this application provides a machine-readable storage medium storing instructions that cause a machine to perform the methods described above.

[0029] The above technical solution firstly utilizes formatting and pre-cleaning judgment strategies to achieve standardized cleaning of original process parameters and automatic filtering of invalid data from non-production stages. This ensures that subsequent prediction calculations are triggered only when the reduction furnace has actually completed cleaning and entered the preparation stage, effectively avoiding noise interference and saving computing resources. Secondly, a parallel and complementary dual prediction mechanism is constructed. On the one hand, a duration prediction model (including structures such as long short-term memory networks) is used to deeply mine the nonlinear characteristics and long-term dependencies of multidimensional process parameters over time, capturing dynamic trends under complex operating conditions. On the other hand, a historical data matching method based on process step mapping is used to extract robust statistical regularities from massive historical production data, providing "experience" references that conform to the physical process logic for prediction. Finally, by fusing the first and second prediction durations through evaluation based on historical databases, the sensitivity of deep learning and the stability of historical statistics can be fully combined, effectively avoiding the bias or overfitting problems that may be caused by a single prediction method. This achieves adaptive and high-precision prediction of the reduction furnace start-up time, significantly reducing the randomness and lag of manual experience judgment.

[0030] Other features and advantages of the embodiments of this application will be described in detail in the following detailed description section. Attached Figure Description

[0031] The accompanying drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the following detailed description to explain the embodiments of this application, but do not constitute a limitation on the embodiments of this application. In the drawings: Figure 1 A flowchart illustrating a method for predicting the start-up time of a reduction furnace according to an embodiment of this application is shown schematically. Figure 2 A flowchart illustrating a method for predicting the start-up time of a reduction furnace according to another embodiment of this application is shown schematically. Figure 3 This illustration schematically shows a logic diagram for determining the start procedure according to an embodiment of this application; Figure 4 This illustration schematically shows a current process step determination flowchart according to an embodiment of the present application; Figure 5 The schematic diagram illustrates a structural diagram of a duration prediction model according to an embodiment of this application; Figure 6 The schematic diagram illustrates a structural diagram of a computer device according to an embodiment of this application. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only for illustration and explanation of the embodiments of this application and are not intended to limit the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0033] It should be noted that if the embodiments of this application involve directional indicators (such as up, down, left, right, front, back, etc.), the directional indicators are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indicators will also change accordingly.

[0034] Furthermore, if the embodiments of this application involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, features defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions of various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed in this application.

[0035] The acquisition, transmission, storage, use, and processing of data in this application comply with relevant laws and regulations. Furthermore, it should be noted that certain software, components, models, and other existing industry solutions may be mentioned in the embodiments of this application. These should be considered exemplary, intended only to illustrate the feasibility of implementing the technical solution of this application, and do not imply that the applicant has already used or necessarily used such solutions.

[0036] It should be noted that all data involved in this application (including but not limited to data used for analysis, data stored, data displayed, etc.) are information and data authorized by the client or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0037] Start-up time refers to the time it takes for a reduction furnace to become ready for start-up after shutdown, cleaning, inspection, airtightness testing, and gas replacement.

[0038] Figure 1 A flowchart illustrating a method for predicting the start-up time of a reduction furnace according to an embodiment of this application is shown schematically. Figure 1 As shown in the figure, this application provides a method for predicting the start-up time of a reduction furnace, which may include the following steps.

[0039] Step 101: Obtain the process parameter timing data and historical database of the target reduction furnace.

[0040] Step 102: Format the time series data of process parameters and output the time series matrix and the current feature vector.

[0041] In this embodiment, the target reduction furnace refers to the polysilicon production equipment for which the start-up conditions need to be predicted. The start-up time refers to the time from the current state of the target reduction furnace to when it meets the start-up conditions. Process parameter time-series data refers to the set of real-time sensor detection parameters obtained from the control system, including time-series information and furnace batch information, covering key production indicators such as reduction furnace pressure, hydrogen flow rate, trichlorosilane flow rate, furnace water and chassis water flow rate, temperature, and the opening degree of various regulating valves. The historical database can refer to a data set storing complete records of past production batches of the target reduction furnace, used for subsequent historical data matching and statistical analysis. Formatting processing refers to the process of cleaning and standardizing the collected raw process parameter time-series data. Specifically, this may include unifying the data to a preset time granularity through downsampling, filling missing data during transmission using linear interpolation, removing abnormal data jump noise using a sliding window filtering algorithm, and normalizing data values ​​to a unified range. The time-series matrix refers to a multi-dimensional data structure constructed based on preprocessed data according to a set historical time window length, containing parameter change trend information from the initial time to the current time, used as input for subsequent deep learning models. The current feature vector is a vector composed of the specific values ​​of each process parameter extracted at the current moment. It reflects the instantaneous operating condition of the equipment and is used for subsequent logical strategy judgment. Through the above steps, noise and errors in the original data can be effectively eliminated, and standard and suitable high-quality input data can be provided for different prediction paths, thus laying a solid data foundation for improving the accuracy of predicting the start-up time of the reduction furnace.

[0042] Step 103: Determine whether the target reduction furnace has completed pre-cleaning based on the current feature vector and the preset pre-cleaning judgment strategy.

[0043] Step 104: If the pre-cleaning has been completed, input the time series matrix and the preset parameter set into the pre-built duration prediction model and output the first predicted duration.

[0044] In this embodiment, the pre-set pre-cleaning judgment strategy refers to a pre-defined set of logical verification rules used to verify whether the equipment meets the conditions for starting the prediction program. Specifically, it includes target values ​​for key process indicators of the furnace water system and various gas valves, as well as a set of allowable fluctuation error thresholds, aiming to accurately identify whether the equipment has completed the basic water supply, venting, and nitrogen replacement processes. Pre-cleaning refers to the initial process steps that the reduction furnace must perform before formally entering the preparation stage for start-up. It mainly covers cleaning the equipment pipelines and replacing the atmosphere inside the furnace to ensure that the production environment meets safety standards. The pre-set parameter set refers to a fixed set of parameters determined through training with historical data for model inference calculations. Specifically, it covers the weight matrices and bias vectors required for the forget gate, input gate, candidate state, and output gate in the long short-term memory model, as well as the weights and bias parameters corresponding to the attention layer. The pre-built duration prediction model refers to a neural network architecture built based on deep learning algorithms, which integrates a long short-term memory model capable of capturing long-term sequence dependency features, a Dropout layer to prevent overfitting, and a fully connected layer for outputting the final result. The first prediction duration refers to the estimated remaining time for the target reduction furnace to reach a state ready for start-up, output by the model after complex nonlinear calculations based on the historical trend characteristics inherent in the input time series matrix. Through the above steps, the system can automatically filter out data interference from non-production preparation stages, triggering deep learning prediction only after cleaning is completed, thereby effectively utilizing data features and significantly improving the accuracy of predicting the start-up time of the reduction furnace.

[0045] Step 105: Determine the current process step of the target reduction furnace based on the current feature vector and the preset start-up judgment strategy; the preset start-up judgment strategy stores the mapping relationship between feature vectors and process steps.

[0046] In this embodiment, the preset start-up judgment strategy refers to a predefined set of logical rules used to identify the specific process node where the equipment is currently operating. It stores the corresponding logic between different combinations of process parameter values ​​and specific equipment operating states, specifically covering the judgment criteria for each pressure holding stage in the mixed gas tightness detection process and each charging / discharging cycle stage in the hydrogen replacement process. The current process step refers to the specific operational stage the target reduction furnace is in during the preparation process for start-up, such as a specific step in the mixed gas tightness detection or a specific number of cycles in hydrogen replacement. By matching the real-time parameter values ​​in the current feature vector with the mapping relationship in the strategy, the system can automatically locate the current operating progress of the equipment from a data perspective. This step avoids the randomness and lag of manual judgment, ensuring that the system can accurately identify the physical state of the equipment, thereby providing a reliable basis for auxiliary prediction based on historical statistical patterns, and thus improving the accuracy of predicting the start-up time of the reduction furnace.

[0047] Step 106: Based on the current process steps, match the second predicted time required for the target reduction furnace to reach the start-up state from the historical database.

[0048] In this embodiment, the "ready-to-start" state refers to the standby state where the reduction furnace has completed a series of preparatory processes, including water supply and venting, gas replacement, and pressure testing, and all process indicators meet production requirements, making it ready to be powered on and start the chemical vapor deposition reaction at any time. The second prediction time refers to a time prediction value calculated based on statistical principles. Specifically, it is obtained by retrieving all past production batches at the same current process step from the historical database, extracting the actual time taken for these batches to reach the "ready-to-start" state from that step, and calculating the arithmetic mean. This step utilizes the statistical patterns of massive historical data to reflect the objective physical trends of equipment operation, effectively smoothing out random fluctuations in a single production run, and providing robust reference data based on empirical rules for the final prediction, thereby significantly improving the accuracy of predicting the reduction furnace's "ready-to-start" time.

[0049] Step 107: Determine the final start time based on the historical database, the first prediction duration, and the second prediction duration.

[0050] In this embodiment, the final start-up time refers to the final time prediction result obtained by comprehensively calculating the first and second prediction times using a weighted fusion algorithm. This combines the nonlinear fitting ability of deep learning models with the robustness of historical statistical patterns, representing the system's optimal estimate of the time when the equipment reaches the start-up standard. In this process, "based on the historical database" means using the actual time consumption data of recent historical production batches stored in the database and the corresponding historical prediction values ​​to calculate the average absolute percentage error of the two prediction methods, and generating a dynamic weight coefficient inversely proportional to the error to ensure that the weight allocation reflects the current actual performance of the model. Through this step, the system can dynamically adjust the proportion of each prediction method in the final result based on its recent performance, effectively avoiding the bias risk of a single prediction model, achieving complementary advantages, and thus significantly improving the accuracy of predicting the start-up time of the reduction furnace.

[0051] The above technical solution firstly utilizes formatting and pre-cleaning judgment strategies to achieve standardized cleaning of original process parameters and automatic filtering of invalid data from non-production stages. This ensures that subsequent prediction calculations are triggered only when the reduction furnace has actually completed cleaning and entered the preparation stage, effectively avoiding noise interference and saving computing resources. Secondly, a parallel and complementary dual prediction mechanism is constructed. On the one hand, a duration prediction model (including structures such as long short-term memory networks) is used to deeply mine the nonlinear characteristics and long-term dependencies of multidimensional process parameters over time, capturing dynamic trends under complex operating conditions. On the other hand, a historical data matching method based on process step mapping is used to extract robust statistical regularities from massive historical production data, providing "experience" references that conform to the physical process logic for prediction. Finally, by fusing the first and second prediction durations through evaluation based on historical databases, the sensitivity of deep learning and the stability of historical statistics can be fully combined, effectively avoiding the bias or overfitting problems that may be caused by a single prediction method. This achieves adaptive and high-precision prediction of the reduction furnace start-up time, significantly reducing the randomness and lag of manual experience judgment.

[0052] In this embodiment, the current feature vector includes the following vector elements: the current ratio of the furnace water regulating valve, the current temperature and current flow rate of the furnace water return, the current opening degree of the hydrogen regulating valve, the current opening degree of the trichlorosilane regulating valve, the current opening degree of the nitrogen regulating valve, and the current pressure inside the target reduction furnace; the preset pre-cleaning judgment strategy includes: a preset error threshold set, the target ratio of the furnace water regulating valve, the target temperature and target flow rate of the furnace water return, the target opening degree of the hydrogen regulating valve, the target opening degree of the trichlorosilane regulating valve, the target opening degree of the nitrogen regulating valve, and the target pressure inside the target reduction furnace; the preset error threshold set includes a first error threshold, a second error threshold, and a third error threshold.

[0053] In this embodiment, the current ratio of the furnace water regulating valve, the current temperature of the furnace water return water, and the current flow rate refer to physical parameters reflecting the real-time operating conditions of the reduction furnace cooling water circulation system, used to characterize the heat exchange status and pipeline patency of the equipment during the preparation stage. The current opening degree of the hydrogen regulating valve, the current opening degree of the trichlorosilane regulating valve, the current opening degree of the nitrogen regulating valve, and the current pressure inside the target reduction furnace refer to physical parameters reflecting the control status of the gas inlet and outlet pipelines of the reduction furnace and the internal gas pressure environment of the furnace, used to characterize the medium flow and pressure maintenance during gas replacement and sealing detection. The first error threshold, the second error threshold, and the third error threshold refer to the maximum allowable deviation range from the target value set for the water system parameters, gas valve opening parameters, and furnace pressure parameters, respectively, used to accommodate sensor measurement errors and minor fluctuations allowed by the process. The target proportions of the furnace water regulating valve, the target temperature and flow rate of the furnace water return, the target opening degree of the hydrogen regulating valve, the target opening degree of the trichlorosilane regulating valve, the target opening degree of the nitrogen regulating valve, and the target pressure inside the reduction furnace are all set according to standard production process specifications. These represent ideal benchmark values ​​indicating that the reduction furnace has completed water supply, venting, and gas replacement operations and reached a safe standby state. By finely setting the above-mentioned multi-dimensional physical parameters and their corresponding stringent judgment criteria, the system can comprehensively quantify and evaluate the cleanliness of the equipment from both fluid control and pressure environment dimensions. This ensures the validity of data input and the consistency of operating conditions, thereby providing a clean starting state basis for subsequent models and significantly improving the accuracy of predicting the start-up time of the reduction furnace.

[0054] In this embodiment of the application, the step of determining whether the target reduction furnace has completed pre-cleaning based on the current feature vector and the preset pre-cleaning judgment strategy includes: determining the status of the water supply and exhaust process of the target reduction furnace based on the current feature vector and the preset pre-cleaning judgment strategy; determining the status of the nitrogen replacement process of the target reduction furnace based on the current feature vector and the preset pre-cleaning judgment strategy; and determining that the target reduction furnace has completed pre-cleaning when both the water supply and exhaust process and the nitrogen replacement process are in a completed state.

[0055] In this embodiment, the status of the water supply and exhaust process refers to the comparison between real-time monitored parameters of the furnace water regulating valve, temperature, and flow rate and preset thresholds, indicating whether the reduction furnace cooling water circulation system has reached a ready state to support subsequent heating operations. The criteria for this determination include the stability of the medium flow and the compliance of heat exchange capacity. The status of the nitrogen replacement process refers to the comparison between the timing changes of the gas regulating valve opening and the furnace pressure response and preset logic, indicating whether the air and impurity gases in the furnace have been completely replaced by nitrogen and the inert atmosphere requirements have been met. The criteria for this determination include the correctness of the valve action sequence and the qualification of sealing and pressure maintenance. The completion status refers to the final determination result where all physical parameter verification conditions in the above specific sub-processes simultaneously satisfy the truth logic. By decoupling the pre-cleaning task into two independent sub-logic branches—liquid pipeline verification and gas pipeline verification—and using logical AND operation as the final judgment condition, a rigorous dual safety and process threshold can be constructed. This completely eliminates the premature start of the prediction program due to misjudgment of a single subsystem, ensuring that the subsequent duration prediction model is built on the physical basis that the equipment's water and gas systems are fully ready, thereby significantly improving the accuracy of predicting the start-up time of the reduction furnace.

[0056] In this embodiment of the application, the step of determining the state of the water supply and exhaust process of the target reduction furnace based on the current feature vector and the preset pre-cleaning judgment strategy includes: determining whether the absolute value of the difference between the current ratio and the target ratio of the furnace water regulating valve is less than a first error threshold; if it is less than the first error threshold, determining whether the absolute value of the difference between the current temperature and the target temperature of the furnace water return is less than or equal to the first error threshold; determining whether the absolute value of the difference between the current flow rate and the target flow rate of the furnace water return is less than or equal to the first error threshold; if all the above conditions are met, the water supply and exhaust process of the target reduction furnace is determined to be in a completed state.

[0057] In this embodiment, determining whether the absolute value of the difference between the current ratio and the target ratio of the furnace water regulating valve is less than the first error threshold means that the system first performs a physical opening verification on the inlet control actuator of the cooling water circulation system. By calculating the deviation between the real-time valve position and the ideal position set by the process, it determines whether the actuator has accurately acted to the preset state that can support the subsequent cleaning flow. Determining whether the absolute value of the difference between the current temperature and the target temperature of the furnace water return is less than or equal to the first error threshold means that after confirming that the valve action is correct, the thermodynamic properties of the cooling medium are further verified. By monitoring the closeness of the return water temperature to the reference temperature, it is ensured that the heat exchange environment inside the furnace has been cooled to a safe operating range. Determining whether the absolute value of the difference between the current flow rate and the target flow rate of the furnace water return is less than or equal to the first error threshold means performing a final check on the fluid dynamic stability of the hydraulic system. By comparing the deviation between the real-time flow rate and the standard flow rate, it is confirmed that the circulation pipeline is unobstructed and the cooling medium supply is sufficient, without air resistance or blockage. If all the above conditions are met, the water supply and venting process of the target reduction furnace is considered complete only when the parameters of the three key physical dimensions—actuator position, medium temperature, and medium flow rate—logically pass stringent tolerance checks. Only then will the system ultimately generate a logical truth signal indicating that the water system is ready. It should be noted that, provided the absolute value of the difference between the current and target ratios of the furnace water regulating valve is less than the first error threshold, the subsequent two judgment conditions can be executed simultaneously, or the former can be executed first, followed by the latter, or vice versa. By employing this logically dependent serial verification mechanism, the requirements for the order of physical condition activation in actual processes can be strictly simulated, avoiding misjudgments caused by sensor data jumps or false compliance of a single parameter. This ensures that the water system's readiness determination is based on a solid physical foundation, thereby significantly improving the accuracy of predicting the furnace start-up time.

[0058] In this embodiment of the application, the step of determining the state of the nitrogen replacement process of the target reduction furnace based on the current feature vector and the preset pre-cleaning judgment strategy includes: sequentially judging whether the following conditions are met, wherein the step of judging whether the latter condition is met is performed if the former condition is met: determining whether the absolute value of the difference between the current opening degree of the hydrogen regulating valve and the target opening degree of the hydrogen regulating valve is less than or equal to a second error threshold; determining whether the absolute value of the difference between the current opening degree of the trichlorosilane regulating valve and the target opening degree of the trichlorosilane regulating valve is less than or equal to a second error threshold; determining whether the absolute value of the difference between the current opening degree of the nitrogen regulating valve and the target opening degree of the nitrogen regulating valve is less than or equal to a second error threshold; determining whether the absolute value of the difference between the current pressure and the target pressure in the target reduction furnace is less than or equal to a third error threshold; determining whether the latest current opening degree of the nitrogen regulating valve is less than or equal to the second error threshold; if all the above conditions are met, the nitrogen replacement process is determined to be in a completed state.

[0059] In this embodiment, determining whether the absolute value of the difference between the current opening degree and the target opening degree of the hydrogen regulating valve, the absolute value of the difference between the current opening degree and the target opening degree of the trichlorosilane regulating valve, and the absolute value of the difference between the current opening degree and the target opening degree of the nitrogen regulating valve are less than or equal to the second error threshold means that the system, according to a strict process sequence, sequentially verifies whether the actuators of each key gas pipeline have accurately moved to the preset opening position to ensure that the hydrogen and trichlorosilane pipelines are in the correct standby configuration and that the nitrogen pipeline has been opened for charging. Determining whether the absolute value of the difference between the current pressure and the target pressure in the target reduction furnace is less than or equal to the third error threshold means that, after confirming that the gas valve array status is correct, the gas pressure environment inside the furnace is physically verified to ensure that nitrogen has been successfully charged and a specific positive pressure level sufficient to displace residual air has been established. Determining whether the latest current opening degree of the nitrogen regulating valve is less than or equal to the second error threshold refers to the final verification of the nitrogen supply cut-off action after the furnace pressure reaches the target, marking the physical termination of the pressurization operation at this stage. If all the above conditions are met, the nitrogen purging process is considered complete only if the series of actions—valve opening, pressure building, and valve closing—strictly follow the physical causal sequence and all parameters meet the target. This mandatory serial logic verification effectively filters out misjudgments caused by reversed operation sequences or instantaneous fluctuations in a single parameter, ensuring that the duration prediction model is built on a reliable basis that the furnace atmosphere has been physically purged, thereby significantly improving the accuracy of predicting the start-up time of the reduction furnace.

[0060] In this embodiment, the duration prediction model includes a long short-term memory model, a dropout layer, and a fully connected layer. The long short-term memory model includes a forget gate and an input gate. The step of inputting a time series matrix and a preset parameter set into the pre-constructed duration prediction model and outputting a first predicted duration includes: determining the cell state data at the current time based on the time series matrix, the hidden state data and cell state data output by the duration prediction model at the previous time step, and the preset parameter set; determining the context feature vector based on the time series matrix, the hidden state data output by the duration prediction model at the previous time step, the cell state data at the current time step, and the preset parameter set; inputting the context feature vector into the dropout layer to reduce overfitting and outputting the processed feature data; inputting the feature data into the fully connected layer for calculation, and activating the calculation result through the ReLU function to output the first predicted duration.

[0061] In this embodiment, the Long Short-Term Memory (LSTM) model refers to a recurrent neural network architecture that solves long-sequence dependency problems by introducing a gating mechanism. It effectively memorizes and extracts the complex nonlinear features of the reduction furnace process parameters over time. The Dropout layer is a regularized structure that randomly discards some neuron connections during neural network training and inference, aiming to enhance the model's generalization ability and prevent overfitting. The fully connected layer is a linear transformation network layer that maps the extracted high-dimensional abstract features to the final one-dimensional time value. The forget gate and input gate are logical units within the model used to regulate information flow, responsible for filtering historical redundant information to be discarded and current key input information to be stored, respectively. Hidden state data and cell state data are internal vectors generated by the model during time step iterations, representing short-term external output features and long-term internal memory carriers, respectively. The context feature vector is a highly condensed feature expression of key temporal information at the current moment, calculated based on the current input and historical state data through an attention mechanism or other aggregation algorithms. The ReLU function is a modified linear unit activation function used to ensure that the first predicted duration value of the final output is non-negative. By constructing a deep learning model that includes the aforementioned refined components, we can delve into the deep patterns in the time series data of process parameters, accurately capture long-term trends using gating mechanisms, and ensure the robustness and physical rationality of the prediction results through regularization and nonlinear activation processing, thereby significantly improving the accuracy of predicting the start-up time of reduction furnaces.

[0062] In this embodiment, the Long Short-Term Memory (LSTM) model includes a forget gate and an input gate; the preset parameter set includes: preset trainable weight parameters, a preset forget gate bias vector, a preset input gate bias vector, a preset input gate weight matrix, a preset candidate state weight matrix, and a preset candidate state bias vector; the step of determining the cell state data at the current time based on the time series matrix, the hidden state data and cell state data output by the duration prediction model at the previous time step, and the preset parameter set includes: combining the time series matrix, the hidden state data output by the duration prediction model at the previous time step, the preset trainable weight parameters, and the preset forget gate bias vector. Input the input vector to the forget gate and output the forget gate vector data; input the time series matrix, the hidden state data output by the duration prediction model at the previous time step, the preset input gate bias vector, and the preset input gate weight matrix to the input gate and output the input gate vector data; determine the candidate state data based on the time series matrix, the hidden state data output by the duration prediction model at the previous time step, the preset candidate state weight matrix, and the preset candidate state bias vector; determine the cell state data at the current time step based on the candidate state data, the cell state data output by the duration prediction model at the previous time step, the forget gate vector data, and the input gate vector data.

[0063] In this embodiment, the preset trainable weight parameters, preset forget gate bias vector, preset input gate bias vector, preset input gate weight matrix, preset candidate state weight matrix, and preset candidate state bias vector refer to a fixed set of network parameters optimized and determined based on a large amount of historical sample data during the model training phase. As the operational kernel of the neural network, these parameters determine the model's rate of forgetting historical information and its sensitivity to new input information when processing time-series data. The forget gate vector data and input gate vector data refer to gating coefficients between 0 and 1, calculated using a nonlinear activation function. These are used to quantify the proportion of internal memory retained from the previous time step and the weight for accepting new information at the current time step, respectively. The candidate state data refers to temporary memory features containing potential key information at the current time step, calculated by combining the current input with the hidden state from the previous time step. The cell state data at the current time step refers to the core internal memory vector, generated by fusing the filtered historical memory with the weighted new current memory through element-wise operations. This vector carries the long-term evolution law and short-term fluctuation characteristics of the reduction furnace process parameters. The refined state update operation based on the gating mechanism described above can effectively solve the gradient vanishing problem in long sequence training. It can accurately simulate the memory retention and update process of process data at the mathematical level, ensuring that the model can dynamically capture and accumulate key temporal dependency features that affect the start-up time, thereby significantly improving the accuracy of predicting the start-up time of the reduction furnace.

[0064] In this embodiment, the duration prediction model further includes an attention layer; the long short-term memory model further includes an output gate; the preset parameter set includes: preset output gate weights, preset output gate bias parameters, preset attention weight parameter set, and preset attention bias parameters; the step of determining the context feature vector based on the time series matrix, the hidden state data output by the duration prediction model at the previous time step, the cell state data at the current time step, and the preset parameter set includes: inputting the preset output gate weights, preset output gate bias parameters, the time series matrix, and the hidden state data output by the duration prediction model at the previous time step into the output gate to obtain output gate vector data; determining the hidden state data at the current time step based on the output gate vector data and the cell state data at the current time step; inputting the hidden state data at the current time step, the preset attention weight parameter set, and the preset attention bias parameters into the attention layer to output the attention score at the current time step; determining the normalized weight corresponding to the attention score at the current time step based on the sum of the attention scores at the current time step and all historical time steps; and determining the context feature vector based on the normalized weight and the hidden state data at the current time step.

[0065] In this embodiment, the attention layer refers to a neural network mechanism that can automatically learn and identify the contribution of data from different time steps in the input sequence to the prediction result, aiming to simulate the human characteristic of focusing on key information while ignoring irrelevant noise. The output gate is a logical component in the Long Short-Term Memory (LSTM) model used to control the proportion of internal cellular memories at the current moment that can be activated and transmitted to the outside as output features. Preset output gate weights, preset output gate bias parameters, preset attention weight parameter sets, and preset attention bias parameters are fixed mathematical coefficients determined during model training and used to finely adjust the output gate opening and calculate the attention score. The output gate vector data refers to adjustment coefficients between 0 and 1 calculated through the activation function, used to filter cellular states. The hidden state data at the current moment refers to the output vector generated by element-wise multiplication of the output gate vector data and the nonlinearly activated cellular state data, containing both historical long-term dependencies and reflecting the characteristics of the current moment. The attention score at the current moment refers to a scalar value reflecting the importance of the data at the current time step, calculated based on the current hidden state and context information through a nonlinear transformation. Normalized weights refer to the probability distribution values ​​obtained by exponentially operating and normalizing the attention scores, ensuring that the sum of the weights for all time steps is 1. Context feature vectors are comprehensive feature representations generated by weighted summation of the historical hidden state sequence using normalized weights, focusing on high-weight key moment information. By introducing an output gate to selectively output from internal memory and combining this with dynamic weighted focusing of key time step features using an attention layer, noise interference from non-critical time period data can be effectively suppressed, enhancing the model's ability to capture key signs of furnace start-up, thereby significantly improving the accuracy of predicting the furnace start-up time.

[0066] In this embodiment, the historical database includes historical production batch data; matching the target reduction furnace from the current process step to the start-up state from the historical database according to the current process step includes determining the second prediction time according to the following formula:

[0067] in, This is the second prediction duration; This represents the total number of times the current process step appears in historical production batch data. This refers to the actual time interval between the kth occurrence of the current process step in the historical production batch data and the start-up status of the tool. This represents the current process step.

[0068] In one embodiment, historical production batch data may refer to production data from the most recent three months. The actual interval from the start-up state may refer to the time required for the reduction furnace to reach the start-up state (start-up conditions) from the current process step.

[0069] In this embodiment of the application, the step of determining the final start time based on the historical database, the first predicted duration, and the second predicted duration includes: determining a first difference between the first predicted duration and the corresponding actual duration in the historical production batch data; determining a second difference between the second predicted duration and the corresponding actual duration in the historical production batch data; determining a first weight of the first predicted duration and a second weight of the second predicted duration based on the first difference and the second difference; and determining the final start time based on the first weight, the second weight, the first predicted duration, and the second predicted duration.

[0070] In this embodiment, the actual duration refers to the true physical time consumed by the target reduction furnace during past production processes, from a specific process step node to a state where it meets the start-up conditions, as recorded in historical production batch data. The first difference and the second difference are evaluation indicators calculated based on statistical principles, quantifying the average deviation between the prediction results generated by the duration prediction model and the prediction results obtained by matching with the historical database relative to the actual duration. Specifically, they are expressed as the average absolute percentage error within a certain statistical period. The first weight and the second weight are normalization coefficients calculated based on the inverse error principle. Their values ​​are negatively correlated with the corresponding differences, aiming to dynamically allocate their contribution to the final result based on the recent actual performance of the two prediction methods. Determining the final start-up time refers to the calculation process of fusing the predicted durations obtained from the two different mechanisms according to their corresponding weight coefficients using a linear weighted summation algorithm. Through the above steps, the system can construct a hybrid prediction mechanism with adaptive calibration capabilities. It uses the measured feedback from recent historical data to dynamically suppress prediction branches with large errors, and fully combines the ability of deep learning to capture nonlinear trends with the robustness of statistical rules to physical conditions, thereby significantly improving the accuracy of predicting the start-up time of the reduction furnace.

[0071] In this embodiment of the application, determining the first difference between the first predicted duration and the corresponding actual duration in the historical production batch data includes: determining the first difference according to the following formula:

[0072] Determining the second difference between the second predicted duration and the corresponding actual duration in historical production batch data includes: determining the second difference according to the following formula:

[0073] in, This is the first difference; The second difference; n is the total number of batches in the historical production batch data; The first in the historical production batch data Within each batch, the predicted duration generated by the duration prediction model; The first in the historical production batch data Within each batch, the predicted duration is obtained by matching with the historical database; The first in the historical production batch data Within each batch, the actual time interval between the current process step and the start-up state.

[0074] In this embodiment of the application, determining the first weight and the second weight of the first prediction duration based on the first difference and the second difference includes: determining the first weight and the second weight according to the following formula:

[0075]

[0076]

[0077] in, It is the first weight; As the second weight; This is the first difference; This is the second difference.

[0078] In this embodiment of the application, determining the final start time based on the first weight, the second weight, the first prediction duration, and the second prediction duration includes: determining the final start time according to the following formula:

[0079] in, For the final start time; It is the first weight; The first prediction duration; As the second weight; This is the second prediction duration.

[0080] The above technical solution firstly utilizes formatting and pre-cleaning judgment strategies to achieve standardized cleaning of original process parameters and automatic filtering of invalid data from non-production stages. This ensures that subsequent prediction calculations are triggered only when the reduction furnace has actually completed cleaning and entered the preparation stage, effectively avoiding noise interference and saving computing resources. Secondly, a parallel and complementary dual prediction mechanism is constructed. On the one hand, a duration prediction model (including structures such as long short-term memory networks) is used to deeply mine the nonlinear characteristics and long-term dependencies of multidimensional process parameters over time, capturing dynamic trends under complex operating conditions. On the other hand, a historical data matching method based on process step mapping is used to extract robust statistical regularities from massive historical production data, providing "experience" references that conform to the physical process logic for prediction. Finally, by fusing the first and second prediction durations through evaluation based on historical databases, the sensitivity of deep learning and the stability of historical statistics can be fully combined, effectively avoiding the bias or overfitting problems that may be caused by a single prediction method. This achieves adaptive and high-precision prediction of the reduction furnace start-up time, significantly reducing the randomness and lag of manual experience judgment.

[0081] The following is an example of this application: Figure 2 A flowchart illustrating a method for predicting the start-up time of a reduction furnace according to another embodiment of this application is shown schematically. Figure 2 As shown. This embodiment consists of the following 6 steps. Details are as follows: Step 1, Data Acquisition: The pre-start process includes a series of steps such as water purging and venting, nitrogen purging, gas mixing and sealing, hydrogen purging, and waiting for start-up.

[0082] The corresponding real-time sensor parameters include, but are not limited to: reduction furnace pressure data, hydrogen flow rate data, TCS flow rate data, furnace drum water flow rate, chassis water flow rate, furnace drum water temperature, hydrogen regulating valve opening data, nitrogen regulating valve opening data, trichlorosilane regulating valve opening data, transformer pressure, flash tank level, TCS storage tank level, etc.

[0083] All parameters must include timing information and furnace number information, and be arranged according to the acquisition timing sequence.

[0084] Step 2, Data Preprocessing: Data acquisition from databases may encounter problems such as misalignment of data time sequence granularity, data transmission issues, data noise, and data omissions. This invention will preprocess the data in the following ways to provide reliable data for subsequent calculations.

[0085] First, data time-series granularity alignment is performed. Research indicates that data collection typically uses a second-level granularity. To align the granularity, a downsampling method is used to reduce the sampling rate. Data is collected at a minute-based frequency, and the average value of the data collected within one minute is extracted. Here, m represents the amount of data collected in the current minute, a represents the corresponding collected data value, and s represents the final collected data value for the corresponding minute. The granularity alignment formula is as follows: In minutes, This represents sampled data from hour 00 to hour 01; for example, reduction furnace pressure data. This is the average value from 00 minutes of the xth hour to 01 minutes of the next xth hour during data collection.

[0086] Since only the data trend characteristics need to be judged, the accuracy requirement is only to represent the trend. Linear interpolation is used to fill in the missing data. y represents the missing value, and x represents the sequence number of the missing value. The missing value is the value preceding the missing value. for Corresponding serial number, The value that follows the missing value. for The corresponding sequence number. When m=0, the linear interpolation formula is as follows:

[0087] Using sliding window filtering to remove data jumps: Since there may be data jumps in the transmitted data, which may affect subsequent training, sliding window filtering is used to remove data jumps.

[0088] Normalize the data. It is the minimum value of the entire dataset. It is the maximum value of the entire dataset. These are the values ​​that need to be normalized. This is the current normalized result, rounded to four decimal places. The normalization formula is as follows:

[0089] Data formatting: Initialize the LSTM model data according to the content required by the training set.

[0090] For the LSTM model, the dataset has the following requirements: Assume the sensor data is converted into a time series matrix. Where T is the total time step count (if one minute is considered one step, the total time step count is the total number of minutes), and F is the number of features (number of sensors or number of abstract features, such as: Tcs flow rate, hydrogen flow rate, furnace water flow rate, reduction furnace pressure, etc.). The corresponding window parameters are as follows: minimum window length. Maximum window length Time step index Generate a historical data training window from the initial time to the indexed time t according to the time step and time step index:

[0091] in: Let F be the sensor at time t. The sensor data are in the following order: reduction furnace pressure, hydrogen flow rate, nitrogen flow rate, TCS flow rate, furnace water flow rate, transformer power, flash tank level, and TCS storage tank level. This represents the data window from the initial time point to time point t.

[0092] Step 3, Startup judgment of the program: Figure 3 This diagram schematically illustrates a logic diagram for determining a start-up procedure according to an embodiment of this application. For example... Figure 3 As shown, the start-up time prediction program will only proceed if the characteristics of ventilation and drainage and nitrogen replacement are met simultaneously. This represents the allowable error value for each formula. All settings can be adjusted independently according to production conditions. The constants in the formulas are production settings and can be adjusted later based on production conditions. Water flow and air venting judgment formula: , This represents the current boiler drum water regulating valve ratio. This represents the target proportion of water in the furnace drum, which can be adjusted according to production conditions, and is expressed as a percentage (%). t represents the current boiler drum water return temperature. This represents the target return water temperature of the furnace drum, which is adjustable according to production conditions, and is expressed in degrees Celsius (°C). , This represents the current return water flow rate of the boiler drum. This represents the target flow rate of water in the furnace drum, which is adjustable according to production conditions, and is expressed in tons per hour (t / h). The general formula is: When the result is 1, the reduction furnace passes the ventilation and drainage status test; when the result is 0, it fails and needs to be tested again until it passes.

[0093] Nitrogen replacement judgment formula or logic: N represents the opening degree of the nitrogen regulating valve, H represents the opening degree of the hydrogen regulating valve, Tcs represents the opening degree of the trichlorosilane regulating valve, in percentage. P represents the furnace pressure, in MPa. To accommodate acceptable errors in each process step, all constants and errors can be varied with process adjustments. This state is for replacing the air composition within the reduction furnace.

[0094] Check if the hydrogen regulating valve is open to 50%.

[0095] Check whether the trichlorosilane regulating valve is opened to 50%.

[0096] Check whether S1 and S2 are met, and adjust the opening of the nitrogen regulating valve to 50%.

[0097] Does it meet S3, where the furnace pressure reaches 0.4 MPa?

[0098] Check if S4 is met and if the nitrogen regulating valve is closed.

[0099] Does the entire process meet the requirements?

[0100] When both conditions are met and At this time, the current reduction furnace enters the next program through the start-up judgment program of the start-up program.

[0101] Step 4, Status Judgment (Current Process Step Judgment): Figure 4 This schematically illustrates a current process step determination flowchart according to an embodiment of this application. For example... Figure 4 As shown. The formula for judging the tightness of the mixed gas is: N represents the opening degree of the nitrogen regulating valve, H represents the opening degree of the hydrogen regulating valve, Tcs represents the opening degree of the TCS (trichlorosilane) regulating valve, in percentage. P represents the furnace pressure, in megapascals. To account for acceptable errors in each process step, all constants and errors can be varied with process adjustments. The purpose of this step is to check whether the airtightness of the reduction furnace meets production standards.

[0102] Check whether the nitrogen regulating valve opening reaches 50%.

[0103] Does it meet S1? Is the pressure inside the furnace around 0.3 MPa?

[0104] Check if S2 is met and if the nitrogen regulating valve is closed.

[0105] Does it meet S3? Does the hydrogen regulating valve reach 3%?

[0106] Does it meet S4? Does the pressure inside the furnace reach about 0.6 MPa?

[0107] Does it meet S5? Does the furnace pressure reach 0.05 MPa?

[0108] Does it meet the full process requirements?

[0109] Once the entire process is completed, the mixed airtightness is successfully determined, and the process proceeds to the next step.

[0110] Hydrogen replacement judgment formula: N represents the opening degree of the nitrogen regulating valve, H represents the opening degree of the hydrogen regulating valve, Tcs represents the opening degree of the TCS regulating valve, and the unit is percentage. P represents the furnace pressure, and the unit is megapascals. To account for acceptable errors in each step, all constants and errors can be varied with process adjustments. The purpose of this process is to displace nitrogen and other impurities from the reduction furnace.

[0111] Check if the hydrogen regulating valve is adjusted to around 3%.

[0112] Check if the pressure inside the furnace has been increased to around 0.3 MPa.

[0113] Check if the pressure inside the furnace has dropped to around 0.05 MPa.

[0114] The first replacement.

[0115] , the second substitution.

[0116] The third replacement.

[0117] The fourth replacement.

[0118] After four replacement conditions are met, the hydrogen replacement is complete.

[0119] Start-up waiting state: After the above process is completed, hydrogen gas is introduced for pressurization. Once the pressure inside the furnace reaches the sealing pressure, it automatically enters the start-up waiting state. The current state is start-up complete, waiting for the reduction and start-up command.

[0120] The above steps can be used to obtain the various states of the reduction furnace data, where... The time required for determining the airtightness of the mixture is divided into 6 steps, each with a different time to reach the start-up state. This is written as a formula and will not be elaborated further. The hydrogen replacement time is divided into 4 steps, and the time to reach the start-up state is different for each step. It is written as a formula here and will not be elaborated further.

[0121]

[0122] Where i is the i-th data. It represents the time from the state corresponding to the i-th data (current process step) to the start state.

[0123] pass Thus we obtain , This represents the average time from the current process step to the start-up state for a single batch, where... This represents the total number of times the corresponding status data is recorded for the corresponding batch.

[0124] Based on the above state judgment formula, the current state of the reduction furnace is judged. After successful judgment, the average time to reach the start-up state is calculated based on the corresponding historical data. At the same time, the current state and the average time are output to the subsequent program.

[0125] Step 5: Training the duration prediction model: Figure 5 A schematic diagram illustrating the structure of a duration prediction model according to an embodiment of this application is shown. Figure 5 As shown. For the LSTM model, the overall structure is as follows. Figure 5 The explanation is that while LSTM models support training on dynamically sized data, the tensor shape needs to be fixed in order to accelerate training using GPUs.

[0126] Using the batch-grouped dataset, the training data is grouped by sequence length (e.g., data with similar lengths are grouped together). Within each group, the dataset is padded with zeros at the end according to the maximum time step within the group, aligning data from different time windows to the same dimension. The training set, validation set, and test set are divided according to time (using different batches of data as the training set to avoid future data leakage), with a split ratio of 8:1:1.

[0127] The data is input into the LSTM layer, then enters the forget gate to retain useful old memories, as shown in the following formula:

[0128] in Represents the forget gate vector data, This is the state data of the previous hidden layer. For trainable weight parameters, For bias parameters, The Sigmoid function outputs a gating value between 0 and 1.

[0129] The input gate determines which new data to update, as shown in the following formula:

[0130]

[0131] in, The input gate weight matrix, This is the input gate bias vector. The candidate state weight matrix, This is the candidate state bias vector.

[0132] Next, the cell state data is updated using the following formula:

[0133] in, The data after cell update is obtained by adding the forgotten part of the data to the updated part of the candidate data.

[0134] Enter the output gate to generate the current hidden state, using the following formula:

[0135]

[0136] in This is the activation value of the output gate. and These are the weights and bias parameters of the output gate. It is in a hidden state.

[0137] After the data exits the LSTM layer, it enters the attention layer, where weights are assigned to different time steps to focus the model on key time steps. This process consists of three parts. The first step calculates the attention score, using the following formula:

[0138] Where s is the learnable context vector, For weight parameters, For bias parameters, This is the attention score. The principle is to... and Projected onto the same space, through a nonlinear activation function Capture complex interactions to ultimately obtain a scalar score. Reflecting time steps The importance of.

[0139] Next, the attention scores are converted into probability distributions. To ensure the total weights are equal to 1, the normalized weight formula is as follows:

[0140] Next, we will discuss the probability corresponding to the time step. and current hidden layer data Perform a weighted sum to generate a context vector.

[0141]

[0142] Weighted summation It integrates information from all time steps in the sequence, but focuses on retaining the features of high-weight time steps.

[0143] Next, we enter the Dropout layer, where neurons are randomly dropped during training to reduce the probability of overfitting. Then, we enter the fully connected layer. After several fully connected layers, the last fully connected layer has one neuron, which outputs the predicted time. Since we need to estimate the arrival time of the furnace, we need to output a time prediction value. To avoid negative output values, we use the ReLU function to ensure that the output is ≥0 (time cannot be negative). Finally, we obtain the estimated time for the furnace to reach the furnace state.

[0144] After the model is built, it is trained in batches according to the training parameters planned based on the sequence length. A loss function is constructed, and the gradients of each weight of the fully connected layer, attention layer and LSTM layer are calculated based on the loss function. The Adam optimizer is used to update each trainable parameter, and the model is trained again until the model performance reaches the accuracy that meets the conditions of on-site production.

[0145] Step 6, Model Integration: To solve practical problems, the model should be embedded into the program.

[0146] like Figure 2 As shown, the model is placed in the corresponding position, and the preprocessed data after the start-up procedure is successfully determined is received. The LSTM model predicts a time value, and the start-up process determines another time value. A weighted fusion method is used to obtain the final predicted value. The final predicted value = (LSTM weights * LSTM predicted value) + (state judgment weights * state judgment predicted value). The weights of each parameter are calculated using their respective average percentage errors and the overall average percentage error. The specific formula is as follows: Mean absolute percentage error of LSTM:

[0147] Mean absolute percentage error in state judgment:

[0148] The error weighting is calculated inversely as follows: LSTM weights: State judgment weights: ; Weights satisfy constraints =1, ensuring that the predicted values ​​are free from scaling bias.

[0149] Final prediction formula: .

[0150] Where n is the total number of production batches in the past 3 months. This represents the actual time from the point corresponding to the current process time to the start time of the i-th batch. This represents the prediction duration of the LSTM model for the i-th batch. This represents the predicted duration for the time status judgment of the i-th batch. This represents the LSTM prediction time for the current batch. The prediction duration represents the status of the current batch.

[0151] The final estimated arrival time is as follows. The system will determine the current status and return the results to the database.

[0152] Figure 6 A schematic diagram illustrating the structure of a computer device according to an embodiment of this application is provided. Figure 6 As shown. This application provides a computer device that may include: a memory 610 configured to store instructions; and a processor 620 configured to retrieve instructions from the memory 610 and, when executing the instructions, to implement the methods described above.

[0153] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0154] This application also provides a machine-readable storage medium storing instructions that cause a machine to perform the above-described method.

[0155] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0156] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0157] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0158] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0159] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0160] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0161] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0162] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0163] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for predicting the start-up time of a reduction furnace, characterized in that, The method includes: Obtain the timing data and historical database of the process parameters of the target reduction furnace; The process parameter time series data is formatted to output a time series matrix and a current feature vector; Based on the current feature vector and the preset pre-cleaning judgment strategy, determine whether the target reduction furnace has completed pre-cleaning; If the pre-cleaning has been completed, the time series matrix and the preset parameter set are input into the pre-constructed duration prediction model, and the first predicted duration is output. The current process step of the target reduction furnace is determined based on the current feature vector and the preset start-up judgment strategy; the preset start-up judgment strategy stores the mapping relationship between the feature vector and the process step; Based on the current process step, a second predicted time is determined from the historical database that the target reduction furnace needs to transition from the current process step to the start-up state. The final start time is determined based on the historical database, the first predicted duration, and the second predicted duration.

2. The method according to claim 1, characterized in that, The current feature vector includes the following vector elements: the current ratio of the furnace water regulating valve, the current temperature and current flow rate of the furnace water return, the current opening degree of the hydrogen regulating valve, the current opening degree of the trichlorosilane regulating valve, the current opening degree of the nitrogen regulating valve, and the current pressure inside the target reduction furnace. The preset pre-cleaning judgment strategy includes: a preset error threshold set, the target ratio of the furnace water regulating valve, the target temperature and target flow rate of the furnace water return water, the target opening degree of the hydrogen regulating valve, the target opening degree of the trichlorosilane regulating valve, the target opening degree of the nitrogen regulating valve, and the target pressure inside the target reduction furnace. The preset error threshold set includes a first error threshold, a second error threshold, and a third error threshold.

3. The method according to claim 2, characterized in that, The step of determining whether the target reduction furnace has completed pre-cleaning based on the current feature vector and the preset pre-cleaning judgment strategy includes: The state of the water supply and exhaust process of the target reduction furnace is determined based on the current feature vector and the preset pre-cleaning judgment strategy. The state of the nitrogen replacement process of the target reduction furnace is determined based on the current feature vector and the preset pre-cleaning judgment strategy. If both the water supply and exhaust process and the nitrogen replacement process are completed, it is determined that the target reduction furnace has completed the pre-cleaning.

4. The method according to claim 3, characterized in that, The step of determining the state of the water supply and exhaust process of the target reduction furnace based on the current feature vector and the preset pre-cleaning judgment strategy includes: Determine whether the absolute value of the difference between the current ratio and the target ratio of the boiler drum water regulating valve is less than the first error threshold. If it is less than the first error threshold, then determine whether the absolute value of the difference between the current temperature of the boiler drum water return water and the target temperature is less than or equal to the first error threshold. Determine whether the absolute value of the difference between the current flow rate of the boiler drum water return and the target flow rate is less than or equal to the first error threshold. If all the above conditions are met, the water supply and exhaust process of the target reduction furnace is determined to be complete.

5. The method according to any one of claims 2 to 4, characterized in that, The step of determining the state of the nitrogen replacement process of the target reduction furnace based on the current feature vector and the preset pre-cleaning judgment strategy includes: The following conditions are checked sequentially, with the step of checking whether a later condition is met performed only if a previous condition is met: Determine whether the absolute value of the difference between the current opening degree of the hydrogen regulating valve and the target opening degree of the hydrogen regulating valve is less than or equal to the second error threshold; Determine whether the absolute value of the difference between the current opening degree of the trichlorosilane control valve and the target opening degree of the trichlorosilane control valve is less than or equal to the second error threshold. Determine whether the absolute value of the difference between the current opening degree of the nitrogen regulating valve and the target opening degree of the nitrogen regulating valve is less than or equal to the second error threshold; Determine whether the absolute value of the difference between the current pressure in the target reduction furnace and the target pressure is less than or equal to the third error threshold. Determine whether the latest current opening degree of the nitrogen regulating valve is less than or equal to the second error threshold; If all the above conditions are met, the nitrogen replacement process is determined to be complete.

6. The method according to any one of claims 2 to 4, characterized in that, The duration prediction model includes a long short-term memory model, a dropout layer, and a fully connected layer; the long short-term memory model includes a forget gate and an input gate. The step of inputting the time series matrix and the preset parameter set into the pre-constructed duration prediction model and outputting the first predicted duration includes: Based on the time series matrix, the hidden state data and cell state data output by the duration prediction model at the previous time step, and the preset parameter set, determine the cell state data at the current time step. The context feature vector is determined based on the time series matrix, the hidden state data output by the duration prediction model at the previous time step, the cell state data at the current time step, and the preset parameter set. The context feature vector is input into the Dropout layer to reduce overfitting, and the processed feature data is output. The feature data is input into a fully connected layer for calculation, and the calculation result is activated by the ReLU function to output the first prediction duration.

7. The method according to claim 6, characterized in that, The long short-term memory model includes a forgetting gate and an input gate; The preset parameter set includes: preset trainable weight parameters, preset forget gate bias vector, preset input gate bias vector, preset input gate weight matrix, preset candidate state weight matrix, and preset candidate state bias vector; The step of determining the cell state data at the current moment based on the time series matrix, the hidden state data and cell state data output by the duration prediction model at the previous time step, and the preset parameter set includes: The time series matrix, the hidden state data output by the duration prediction model at the previous time step, the preset trainable weight parameters, and the preset forget gate bias vector are input into the forget gate, and the forget gate vector data is output. The time series matrix, the hidden state data output by the duration prediction model at the previous time step, the preset input gate bias vector, and the preset input gate weight matrix are input to the input gate, and the input gate vector data is output. Candidate state data is determined based on the time series matrix, the hidden state data output by the duration prediction model at the previous time step, the preset candidate state weight matrix, and the preset candidate state bias vector. The cell state data at the current moment is determined based on the candidate state data, the cell state data output by the duration prediction model at the previous time step, the forgetting gate vector data, and the input gate vector data.

8. The method according to claim 6, characterized in that, The duration prediction model further includes an attention layer; the long short-term memory model further includes an output gate; the preset parameter set includes: preset output gate weights, preset output gate bias parameters, preset attention weight parameter set, and preset attention bias parameters; The step of determining the context feature vector based on the time series matrix, the hidden state data output by the duration prediction model at the previous time step, the cell state data at the current time step, and the preset parameter set includes: The preset output gate weights, the preset output gate bias parameters, the time series matrix, and the hidden state data output by the duration prediction model at the previous time step are input into the output gate to obtain the output gate vector data. The hidden state data at the current moment is determined based on the output gate vector data and the cell state data at the current moment; The hidden state data at the current moment, the preset attention weight parameter set, and the preset attention bias parameter are input into the attention layer, and the attention score at the current moment is output. The normalized weight corresponding to the attention score at the current moment is determined based on the sum of the attention scores at all historical moments. The context feature vector is determined based on the normalized weights and the hidden state data at the current time.

9. The method according to any one of claims 2 to 4, characterized in that, The historical database includes historical production batch data; The second predicted time required for the target reduction furnace to reach the start-up state from the current process step, matched from the historical database based on the current process step, includes: The second prediction duration is determined using the following formula: in, This is the second prediction duration; The total number of times the current process step appears in the historical production batch data; The first in the historical production batch data k The actual time interval between the occurrence of the current process step and the start-up state; This refers to the current process step.

10. The method according to claim 9, characterized in that, The step of determining the final start time based on the historical database, the first prediction duration, and the second prediction duration includes: Determine the first difference between the first predicted duration and the corresponding actual duration in the historical production batch data; Determine a second difference between the second predicted duration and the corresponding actual duration in the historical production batch data; The first weight of the first prediction duration and the second weight of the second prediction duration are determined based on the first difference and the second difference. The final start time is determined based on the first weight, the second weight, the first prediction duration, and the second prediction duration.

11. The method according to claim 10, characterized in that, The determination of the first difference between the first predicted duration and the corresponding actual duration in the historical production batch data includes: The first difference is determined according to the following formula: The determination of the second difference between the second predicted duration and the corresponding actual duration in the historical production batch data includes: The second difference is determined according to the following formula: in, This is the first difference; This is the second difference; n The total number of batches in the historical production batch data; The first in the historical production batch data Within each batch, the predicted duration generated by the duration prediction model; The first in the historical production batch data Within each batch, the predicted duration is matched using the historical database; The first in the historical production batch data Within each batch, the actual time interval between the current process step and the start-up state.

12. The method according to claim 10, characterized in that, The step of determining the first weight of the first prediction duration and the second weight of the second prediction duration based on the first difference and the second difference includes: The first weight and the second weight are determined according to the following formula: in, This is the first weight; This is the second weight; This is the first difference; This is the second difference.

13. The method according to claim 10, characterized in that, Determining the final start time based on the first weight, the second weight, the first prediction duration, and the second prediction duration includes: The final start time is determined according to the following formula: in, The final start time; This is the first weight; The first prediction duration; This is the second weight; This is the second prediction duration.

14. A computer device, characterized in that, include: The memory is configured to store instructions; as well as A processor is configured to retrieve the instructions from the memory and, when executing the instructions, to implement the method according to any one of claims 1 to 13.

15. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method according to any one of claims 1 to 13.

16. A machine-readable storage medium, characterized in that, The machine-readable storage medium stores instructions for causing the machine to perform the method according to any one of claims 1 to 13.

Images