Soft measurement model construction method, polycrystalline silicon reduction furnace control method and equipment

Abstract

The invention relates to a polycrystalline silicon reduction furnace soft measurement model construction method, a polycrystalline silicon reduction furnace control method and computer equipment. The method comprises the following steps: determining an input variable set and an output variable set; the output variable set comprises in-furnace definition, and the input variable set comprises hydrogen flow, silicon rod diameter, heating power and cooling water flow; determining a lag time constant set; constructing an in-furnace atmosphere temperature estimation sub-model; constructing a to-be-trained model structure based on the input variable set, the output variable set, the lag time constant set and the in-furnace atmosphere temperature estimation sub-model; training the to-be-trained model structure based on historical operation data of the polycrystalline silicon reduction furnace to obtain a corresponding soft measurement model; the soft measurement model is configured to receive real-time data of the input variable set and output at least an in-furnace sharpness estimate. By adopting the method, the real-time and accurate estimation of the definition in the furnace can be realized, and an accurate state feedback basis is provided for the multivariable cooperative control of the polycrystalline silicon reduction furnace.

Description

Soft measurement model construction method, polysilicon reduction furnace control method and equipment Technical Field

[0001] This application relates to the field of polysilicon production control, and in particular to a method for constructing a soft measurement model of a polysilicon reduction furnace, a method for controlling a polysilicon reduction furnace, and computer equipment. Background Technology

[0002] Polycrystalline silicon is a key basic material for the photovoltaic and semiconductor industries, and its purity and structural integrity directly affect the performance of downstream products. In the reduction furnace production of the modified Siemens process, the silicon rod surface temperature and furnace clarity are the core process conditions that directly determine the deposition quality, production safety, and energy consumption. Among them, furnace clarity reflects the concentration and atomization degree of by-product particles and is a key leading indicator for predicting product contamination risks and preventing safety accidents.

[0003] Currently, soft sensing technology has been applied to the process monitoring of reduction furnaces to address the problem of some parameters being difficult to measure directly. Specifically, by establishing models, measurable parameters are used to estimate and monitor operating indicators such as exhaust gas temperature, operating power consumption, silicon rod diameter, deposition rate, and unit power consumption. This method focuses on assessing production efficiency and energy consumption. However, the soft sensing model proposed by this method has the following limitations: 1. Existing soft sensing models fail to cover furnace clarity, a key controlled variable crucial to the quality and safety of core products. This results in operators lacking real-time, quantitative means to perceive the risk of contamination within the furnace, relying only on delayed manual observation or experience-based judgment, creating blind spots in product quality control and safety early warning.

[0004] 2. Due to the complex influence of ambient temperature, gas flow, and chemical reaction kinetics on furnace clarity, it exhibits significant dynamic hysteresis characteristics. Existing soft sensor models are mostly static or quasi-static models, lacking compensation mechanisms for the system's dynamic hysteresis characteristics, resulting in sluggish estimation responses and insufficient accuracy.

[0005] Due to the lack of accurate dynamic perception of the aforementioned key quality states, existing control systems struggle to achieve precise control. Summary of the Invention

[0006] Therefore, it is necessary to provide a method for constructing a soft measurement model for a polysilicon reduction furnace, a method for controlling a polysilicon reduction furnace, and a computer device to address the aforementioned technical problems.

[0007] In a first aspect, this application provides a method for constructing a soft sensor model for a polysilicon reduction furnace. The method includes: determining an input variable set and an output variable set; the output variable set includes furnace clarity, and the input variable set includes hydrogen flow rate, silicon rod diameter, heating power, and cooling water flow rate; determining a lag time constant set to characterize at least the lag effect of the hydrogen flow rate and / or the silicon rod diameter on the furnace clarity; constructing a furnace atmosphere temperature estimation sub-model to calculate an estimated furnace atmosphere temperature based on the heating power, the cooling water flow rate, and furnace heat loss; constructing a training model structure based on the input variable set, the output variable set, the lag time constant set, and the furnace atmosphere temperature estimation sub-model; training the training model structure based on historical operating data of the polysilicon reduction furnace to obtain a corresponding soft sensor model; the soft sensor model is configured to receive real-time data from the input variable set and output at least an estimated furnace clarity value.

[0008] In one embodiment, the set of output variables further includes the silicon rod surface temperature, and the set of input variables further includes the heating current; the set of hysteresis time constants is also used to characterize the hysteresis effect of the heating current on the silicon rod surface temperature; the soft measurement model is also configured to output an estimate of the silicon rod surface temperature.

[0009] In one embodiment, the set of hysteresis time constants includes a first hysteresis time constant, a second hysteresis time constant, and a third hysteresis time constant; wherein, the first hysteresis time constant is used to characterize the hysteresis time of the effect of heating current change on silicon rod surface temperature; the second hysteresis time constant is used to characterize the hysteresis time of the effect of hydrogen flow rate change on furnace clarity; and the third hysteresis time constant is used to characterize the hysteresis time of the effect of silicon rod diameter change on model coupling relationship.

[0010] In one embodiment, the first lag time constant ranges from 10 to 30 seconds, the second lag time constant ranges from 20 to 40 seconds, and the third lag time constant ranges from 30 to 60 seconds.

[0011] In one embodiment, the furnace atmosphere temperature estimation sub-model calculates the furnace atmosphere temperature estimate based on the following formula. : ;in, The heating power; The cooling water flow rate; The specific heat capacity of water; The temperature difference between the inlet and outlet of the cooling water; This refers to the heat loss of the furnace body; The mass of gas inside the furnace; This represents the specific heat capacity of the gas.

[0012] In one embodiment, the structure of the model to be trained is as follows: ;in, The surface temperature of the silicon rod; For the clarity inside the furnace; The heating current; The hydrogen flow rate; The diameter of the silicon rod; These are the model coefficients; , This is the error term; The first lag time constant; This is the second lag time constant; The third lag time constant is mentioned above.

[0013] In one embodiment, the method further includes: dividing the polysilicon reduction furnace operation process into multiple growth stages corresponding to the silicon rod diameter ranges according to multiple preset silicon rod diameter ranges; and independently training a corresponding soft measurement model for each growth stage.

[0014] In one embodiment, the plurality of growth stages include a first stage, a second stage, and a third stage; the silicon rod diameter range corresponding to the first stage is no greater than 50 mm, the silicon rod diameter range corresponding to the second stage is greater than 50 mm and less than 100 mm, and the silicon rod diameter range corresponding to the third stage is no less than 100 mm.

[0015] Secondly, this application also provides a method for controlling a polysilicon reduction furnace, comprising: acquiring real-time operating parameters of the polysilicon reduction furnace; inputting the real-time operating parameters into a soft measurement model to obtain at least an estimated value of the furnace's internal sharpness; and generating control instructions for the operating variables based at least on a first deviation between the estimated value of the furnace's internal sharpness and a sharpness setpoint.

[0016] In one embodiment, after inputting the real-time operating parameters into the soft sensor model, the method further includes: obtaining the estimated surface temperature of the silicon rod output by the soft sensor model; the step of generating control instructions for the operating variable includes: generating control instructions for the operating variable based on the first deviation and the second deviation between the estimated surface temperature of the silicon rod and the temperature setpoint.

[0017] In one embodiment, a control command for the operating variable is generated based on the first deviation and the second deviation, including: if the absolute value of the second deviation exceeds a temperature threshold, a first control command for the heating current is generated based on the estimated surface temperature of the silicon rod and a temperature setpoint; if the absolute value of the second deviation does not exceed the temperature threshold and the absolute value of the first deviation exceeds a sharpness threshold, a second control command for the hydrogen flow rate and / or trichlorosilane flow rate is generated based on the estimated sharpness in the furnace and a sharpness setpoint.

[0018] In one embodiment, the real-time operating parameters include the current silicon rod diameter, and the method further includes: if the absolute value of the second deviation does not exceed the temperature threshold and the absolute value of the first deviation does not exceed the sharpness threshold, then obtaining a third deviation between the current silicon rod diameter and the target silicon rod diameter; and generating a third control command for the trichlorosilane flow rate based on the third deviation.

[0019] In one embodiment, before inputting the real-time operating parameters into the soft sensor model, the method further includes: determining the current growth stage of the polysilicon reduction furnace operation process based on the current silicon rod diameter; the step of inputting the real-time operating parameters into the soft sensor model includes: inputting the real-time operating parameters into the soft sensor model corresponding to the current growth stage.

[0020] In one embodiment, the method further includes: performing a first verification on the soft measurement model at a first frequency; the first verification is performed based on the furnace sharpness estimate, calibration sharpness, silicon rod surface temperature estimate, and calibration temperature, wherein the calibration sharpness and calibration temperature are obtained based on calibration data consistent with the current operating conditions in a historical database; performing a second verification on the soft measurement model at a second frequency; the second verification is performed based on the furnace sharpness estimate, visual sharpness, silicon rod surface temperature estimate, and visual temperature, wherein the visual sharpness and visual temperature are obtained by processing furnace images; the first frequency is greater than the second frequency.

[0021] In one embodiment, performing the first verification includes: obtaining a first absolute deviation between the furnace sharpness estimate and the calibrated sharpness, and a second absolute deviation between the silicon rod surface temperature estimate and the calibrated temperature; if the first absolute deviation is greater than a first sharpness verification threshold, or the second absolute deviation is greater than a first temperature verification threshold, then the following steps are executed repeatedly until a first stopping condition is met: calling a preset number of recent historical running data to retrain the model structure to be trained to obtain a new soft measurement model; based on the new soft measurement model, re-obtaining the corresponding furnace sharpness estimate and silicon rod surface temperature estimate, and calculating a new first absolute deviation and a new second absolute deviation; wherein the first stopping condition includes: the new first absolute deviation is not greater than the first sharpness verification threshold, and the new second absolute deviation is not greater than the first temperature verification threshold; and the number of times the loop is executed reaches a preset number.

[0022] In one embodiment, performing the second verification includes: acquiring the visual sharpness and the visual temperature based on the furnace images; acquiring a first relative deviation between the furnace sharpness estimate and the visual sharpness, and a second relative deviation between the silicon rod surface temperature estimate and the visual temperature; if the first relative deviation is greater than a second sharpness verification threshold, or the second relative deviation is greater than a second temperature verification threshold, then continuously acquiring multiple new furnace images and correspondingly calculating multiple first relative deviations and second relative deviations; calculating a first average value of the multiple sets of first relative deviations, and calculating a second average value of the multiple sets of second relative deviations; if the first average value is greater than the second sharpness verification threshold, or the second average value is greater than the second temperature verification threshold, then determining that the current soft measurement model is abnormal.

[0023] Thirdly, embodiments of this application provide a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of a method for constructing a soft measurement model of a polysilicon reduction furnace or a method for controlling a polysilicon reduction furnace.

[0024] One of the above technical solutions has the following advantages or beneficial effects: by incorporating furnace clarity into the output variable set and introducing a set of hysteresis time constants to compensate for the dynamic hysteresis effects of key input variables such as hydrogen flow rate and silicon rod diameter, a dynamic soft-sensor model structure including atmosphere temperature estimation is constructed. This method can achieve real-time and accurate estimation of furnace clarity, solving the problem that existing soft-sensor models cannot cover this key controlled variable and do not consider the dynamic characteristics of the system, thus providing an accurate state feedback basis for subsequent multivariate collaborative control. Attached Figure Description

[0025] Figure 1 is a flowchart illustrating a method for constructing a soft measurement model of a polysilicon reduction furnace in one embodiment; Figure 2 is a flowchart illustrating a method for controlling a polysilicon reduction furnace in one embodiment; Figure 3 is an internal structure diagram of a computer device in one embodiment. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0027] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0028] This application addresses the problem that the clarity inside a polysilicon reduction furnace is difficult to measure directly and in real time, and has significant dynamic lag. By constructing a dynamic soft measurement model that includes lag compensation, it achieves high-precision estimation of the clarity inside the furnace and the surface temperature of the silicon rod.

[0029] Based on the estimated values ​​output by this dynamic soft measurement model, a multivariate priority collaborative control and a two-level verification mechanism are implemented to improve the quality control level and operational safety of polysilicon production.

[0030] In one embodiment, as shown in Figure 1, a method for constructing a soft measurement model of a polycrystalline silicon reduction furnace is provided, including the following steps: S102: Determine the input variable set and the output variable set; the output variable set includes furnace clarity, and the input variable set includes hydrogen flow rate, silicon rod diameter, heating power, and cooling water flow rate.

[0031] This step aims to clarify the input and output boundaries of the soft measurement model (hereinafter referred to as the model).

[0032] Among them, the output variable set includes at least furnace clarity; furnace clarity is a key indicator characterizing the visual transparency inside the reduction furnace. Its physical nature reflects the concentration of atomized particles (such as undeposited silicon powder, by-product chloride, etc.) inside the furnace, which directly affects the purity of the deposited silicon rod and production safety. It is a process state that is difficult to measure directly and continuously but is of great importance.

[0033] The input variable set includes a series of process parameters that can be measured online or indirectly obtained and have a significant impact on the output variables. Specifically, the input variable set includes at least: hydrogen flow rate, silicon rod diameter, heating power, and cooling water flow rate. Among them, the silicon rod diameter can be obtained in real time through a visual measurement system installed in the furnace observation window, or estimated based on the deposition time and a preset growth rate model.

[0034] In practical implementation, the process of determining the input and output variable sets may include: based on process mechanism analysis, initially screening potentially relevant variables, and then using historical operating data to perform correlation analysis (such as Pearson correlation coefficient, mutual information analysis, etc.) or feature importance ranking (such as feature importance assessment based on tree models) to quantify the degree of influence of each candidate input variable on the output variable; finally, selecting variables with high correlation, strong measurability, and no redundancy to form the input variable set. In addition, the output variable set can be expanded according to control requirements; for example, it can further include the silicon rod surface temperature. In this case, the input variable set needs to be increased with heating current to construct a multi-output soft-sensor model that can simultaneously estimate temperature and clarity.

[0035] A specific implementation method of correlation analysis is shown in the table below, which clarifies the influence mechanism and quantitative relationship of each candidate input variable on in-furnace sharpness:

[0036] Based on the correlation analysis shown in the table above, hydrogen flow rate, silicon rod diameter, atmosphere temperature, and heating current (which can be reflected by heating power) that directly affect temperature can be identified as key input variables for constructing a soft measurement model of furnace clarity.

[0037] S104: Determine the set of hysteresis time constants to characterize at least the hysteresis effect of hydrogen flow rate and / or silicon rod diameter on furnace clarity.

[0038] Because the reduction furnace involves complex gas flow, heat transfer and chemical reaction processes, it takes a certain amount of time for changes in input variables to be transmitted to output variables and have a significant impact, i.e., there is a dynamic lag effect.

[0039] This step aims to determine a set of lag time constants to compensate for this dynamic characteristic in the model.

[0040] It should be noted that the lag time constant set is used at least to characterize the lag time of the effect of hydrogen flow rate and / or silicon rod diameter on furnace clarity. Its physical meaning can be understood as the average time delay from the adjustment of hydrogen flow rate or the increase of silicon rod diameter to the occurrence of a measurable change in furnace clarity.

[0041] Specific methods for determining at least one lag time constant in the set of lag time constants include, but are not limited to: analyzing the response curve of the output variable after a step change in a certain input variable in historical data or controlled experiments, and defining the time required for the response to reach a steady-state change as the lag time constant of that channel. Alternatively, using historical input-output data, fitting a transfer function model or state-space model with lag, and extracting the lag time parameters of each channel from it. Or, providing an initial estimate based on mechanistic knowledge such as the residence time of gas in the furnace and the heat transfer time constant, and then fine-tuning it through data fitting.

[0042] In a preferred embodiment, the set can be embodied as three hysteresis time constants: used to compensate for the hysteresis effects of heating current on temperature, hydrogen flow rate on clarity, and silicon rod diameter on the overall system coupling relationship.

[0043] S106: Construct a sub-model for estimating the furnace atmosphere temperature, which is used to calculate the estimated value of the furnace atmosphere temperature based on the heating power, cooling water flow rate and furnace heat loss.

[0044] It should be explained that the furnace atmosphere temperature is a key environmental parameter affecting the silicon deposition reaction kinetics and particle behavior. However, due to the high temperature and highly corrosive environment inside the furnace, it is difficult to install sensors for direct and continuous measurement. This step aims to construct a sub-model for estimating the furnace atmosphere temperature based on the energy balance principle, in order to achieve indirect real-time estimation of it.

[0045] The core of this sub-model is to establish a balance between the energy input of the heating system, the energy removed by the cooling system, the heat dissipation of the furnace body, and the heat absorption of the atmosphere. Thus, this sub-model transforms the atmosphere temperature, which is difficult to measure directly, into a state variable that can be calculated using measurable parameters (i.e., heating power and cooling water flow rate) and known physical properties, and incorporates it as a key intermediate variable into the subsequent soft measurement model.

[0046] S108: Construct the model structure to be trained based on the input variable set, output variable set, lag time constant set, and furnace atmosphere temperature estimation sub-model.

[0047] This step integrates the previously determined set of variables, the hysteresis compensation mechanism, and the atmosphere temperature estimation sub-model to form a training model structure with a clear mathematical expression. In the training model structure, each output variable can be expressed as a linear or nonlinear combination of the input variable with hysteresis and the estimated atmosphere temperature.

[0048] For example, when the structure of the model to be trained can simultaneously estimate the silicon rod surface temperature and furnace clarity, the structural framework of the model to be trained can be represented as: ;in, and This can be understood as a functional relationship to be determined, which can be linear weighted sum, neural network, support vector regression, etc., without any restrictions here; The surface temperature of the silicon rod; For clarity inside the furnace; This is the heating current; This refers to the hydrogen flow rate; The diameter of the silicon rod; The estimated value of the furnace atmosphere temperature determined by S106 can be obtained based on parameters such as heating power, cooling water flow rate, and furnace heat loss. , which is the set of time lag constants determined by S104; , This represents the error term. The model structure to be trained explicitly embeds hysteresis compensation terms and atmospheric temperature, laying the foundation for accurately describing the dynamic characteristics of the system.

[0049] S110: The structure of the model to be trained is trained based on the historical operating data of the polysilicon reduction furnace to obtain the corresponding soft measurement model; the soft measurement model is configured to receive real-time data of the input variable set and output at least the furnace sharpness estimate.

[0050] This step uses historical operational data collected from the actual production process to train the training model structure constructed by S108, in order to determine the unknown parameters in the model (such as the coefficients of linear combinations, the weights of neural networks, etc.), thereby obtaining a soft measurement model that can be put into use.

[0051] The soft sensor model is trained based on historical operating data of polysilicon reduction furnaces. The specific implementation process includes: First, data preparation, which involves collecting a large amount of historical data covering different operating conditions and different silicon rod growth stages, potentially exceeding ten thousand sets. This data may include all input variable time series defined in S102, as well as reference values ​​for output variables obtained through high-precision but discontinuous offline measurements (such as infrared temperature measurement at specific times, manual observation records) or special calibration experiments. Then, the data undergoes preprocessing such as cleaning, alignment, and normalization.

[0052] Secondly, the model is trained. Specifically, the preprocessed data can be divided into training set and test set. Using the training set data, optimization algorithms (such as least squares method, gradient descent method, genetic algorithm, etc.) are used to solve the model parameters so that the error (such as mean square error, mean absolute error) between the model output and the reference output is minimized.

[0053] Then, the performance of the trained model is evaluated using test set data to ensure that its estimation accuracy (e.g., temperature error ≤ ±2℃, sharpness error ≤ ±3%) and generalization ability meet the requirements.

[0054] Once the soft measurement model is trained, it can be deployed in the control system to receive real-time data of the input variables and output high-precision furnace clarity estimates and silicon rod surface temperature estimates online.

[0055] In the above embodiments of this application, the executing entity can be a computer device with certain computing capabilities. In the polysilicon reduction furnace soft measurement model construction method of the above embodiments, by incorporating furnace clarity into the output variable set and innovatively introducing a set of lag time constants to compensate for the dynamic lag effect of key input variables, and simultaneously integrating an atmosphere temperature estimation sub-model based on energy balance, a complete dynamic soft measurement framework is constructed. This method achieves real-time, high-precision soft measurement of the core quality and safety indicator of the polysilicon reduction furnace, namely furnace clarity, effectively solving the problem that existing models cannot cover this key variable and do not consider the dynamic characteristics of the system, leading to estimation lag and inaccuracy. This provides an accurate and reliable state-aware foundation for subsequent refined multi-variable collaborative control.

[0056] In one embodiment, the output variable set further includes the silicon rod surface temperature, and the input variable set further includes the heating current; the hysteresis time constant set is also used to characterize the hysteresis effect of the heating current on the silicon rod surface temperature; the soft measurement model is also configured to output an estimate of the silicon rod surface temperature.

[0057] In this embodiment, the silicon rod surface temperature is further included in the output variable set, and the heating current is included in the input variable set. Silicon rod surface temperature is another core parameter controlling the deposition rate and silicon rod quality, while heating current is a key operational variable directly affecting the silicon rod heating power. By constructing a dual-output model that simultaneously outputs temperature and clarity, synchronous high-precision estimation of the two most critical controlled variables is achieved, providing the possibility for comprehensive monitoring and coordinated control.

[0058] It should be noted that the structure of this dual-output model can be coupled (i.e., the two outputs share part of the input or hidden layer) or decoupled, which can be determined through training.

[0059] In one embodiment, the set of hysteresis time constants includes a first hysteresis time constant, a second hysteresis time constant, and a third hysteresis time constant; wherein, the first hysteresis time constant is used to characterize the hysteresis time of the effect of heating current change on silicon rod surface temperature; the second hysteresis time constant is used to characterize the hysteresis time of the effect of hydrogen flow rate change on furnace clarity; and the third hysteresis time constant is used to characterize the hysteresis time of the effect of silicon rod diameter change on model coupling relationship. In some preferred embodiments, the value range of the first hysteresis time constant is 10-30 seconds, the value range of the second hysteresis time constant is 20-40 seconds, and the value range of the third hysteresis time constant is 30-60 seconds.

[0060] The above embodiments further define the set of lag time constants, namely, the set of lag time constants includes a first, a second, and a third lag time constant. Among them, the first lag time constant (as described above) The second hysteresis time constant (as described above) characterizes the inertial time by which the surface temperature of the silicon rod responds after the heating current is adjusted. It is primarily determined by thermal inertia and the characteristics of the current control loop. The third lag time constant (as described above) characterizes the time required for a change in the particle concentration (clarity) within the furnace after adjusting the hydrogen flow rate (purge gas), and is related to the gas replacement rate and particle transport process. This characterizes the overall lag in the impact of the slow-varying perturbation of silicon rod diameter growth on the coupling relationship of the entire system, reflecting the slow drift of process characteristics as the growth process progresses.

[0061] By clearly defining the specific lag channels and the typical value ranges of each lag time constant, the dynamic compensation of the model becomes more targeted and physically interpretable, which helps to improve the estimation accuracy and robustness of the model under varying operating conditions.

[0062] In one embodiment, the furnace atmosphere temperature estimation sub-model calculates the furnace atmosphere temperature estimate based on the following formula. : ;in, This refers to the heating power. This refers to the cooling water flow rate; The specific heat capacity of water; The temperature difference between the inlet and outlet of the cooling water; This is for heat loss from the furnace body; The mass of gas inside the furnace; This represents the specific heat capacity of the gas.

[0063] This embodiment utilizes the energy balance formula, as well as measurable heating power, cooling water flow rate, and calculable furnace heat loss to quantify the heat absorption of the atmosphere, and then inversely calculates the atmosphere temperature.

[0064] In one embodiment, the structure of the model to be trained is as follows: ;in, The surface temperature of the silicon rod; For clarity inside the furnace; This is the heating current; This refers to the hydrogen flow rate; The diameter of the silicon rod; These are the model coefficients; , This is the error term; This is the first time lag constant; This is the second lag time constant; This is the third lag time constant.

[0065] In the dual-output linear model structure proposed in this embodiment These are the model coefficients that need to be determined through training with historical data. , This is to account for the acceptable model error. This specific structure puts the aforementioned abstract model framework into practice, improving the feasibility and repeatability of the solution.

[0066] It should be noted that the linear form of the model structure to be trained is simple to compute and easy to implement, and can provide initialization or benchmark references for more complex nonlinear models (such as neural networks).

[0067] In one embodiment, the method for constructing a soft sensor model for a polycrystalline silicon reduction furnace further includes: dividing the polycrystalline silicon reduction furnace operation process into multiple growth stages corresponding to the silicon rod diameter ranges based on multiple preset silicon rod diameter ranges; and independently training a corresponding soft sensor model for each growth stage. In a specific implementation, the multiple growth stages include a first stage, a second stage, and a third stage; the silicon rod diameter range corresponding to the first stage is no greater than 50 mm, the silicon rod diameter range corresponding to the second stage is greater than 50 mm and less than 100 mm, and the silicon rod diameter range corresponding to the third stage is no less than 100 mm.

[0068] Because the heat capacity, surface area, and convective heat transfer conditions of silicon rods change dramatically from initial deposition to growth into thick rods, it is difficult for a single fixed parameter model to maintain high accuracy throughout the entire process.

[0069] This scheme divides the polysilicon reduction furnace operation process into multiple growth stages based on silicon rod diameter ranges (e.g., ≤50mm, 50-100mm, ≥100mm), and independently trains and configures a soft sensor model for each stage, effectively overcoming the model mismatch problem caused by the time-varying characteristics of the process object. Specifically, the historical data corresponding to the first, second, and third stages are different, and the model coefficients and / or error terms corresponding to the independently trained soft sensor models will also be different. Furthermore, the first, second, and third lag time constants are determined based on fitting historical data, so the lag time constant sets for the three stages may also be different, ultimately resulting in three different soft sensor models.

[0070] During implementation, the system can monitor the silicon rod diameter in real time. When the silicon rod diameter crosses a stage threshold, it automatically switches to the model for the corresponding stage for calculation. This embodiment helps improve the adaptability and estimation stability of the soft measurement model throughout the entire growth cycle, which is key to ensuring control accuracy from start to finish.

[0071] In one embodiment, as shown in FIG2, this application provides a polysilicon reduction furnace control method. Based on the soft measurement model established in the above embodiment, the method can implement closed-loop optimization control, including: S202, obtaining the real-time operating parameters of the polysilicon reduction furnace.

[0072] This step aims to acquire real-time operating parameters of the reduction furnace; these parameters should include at least all input variables defined by the soft-sensor model in the above embodiments, such as heating current, hydrogen flow rate, heating power, cooling water flow rate, and the current silicon rod diameter estimated by a vision system or model. This data can be acquired in real-time through field sensors, instruments, and the control system.

[0073] S204, input the real-time operating parameters into a soft measurement model (the soft measurement model provided in any of the above embodiments) to obtain at least an in-furnace sharpness estimate.

[0074] In this step, the real-time operating parameters acquired in S202 are input into the trained and deployed soft measurement model (if staged modeling is used, the growth stage must first be determined based on the current silicon rod diameter, and the model for the corresponding stage must be called). The model can then calculate and output the furnace sharpness estimate online in real time based on these inputs. If the model is a dual-output model, it can simultaneously output the silicon rod surface temperature estimate.

[0075] S206, at least based on the first deviation between the furnace clarity estimate and the clarity setpoint, generates control instructions for the operating variables.

[0076] The system compares the estimated furnace sharpness output by the model with the preset sharpness setting (e.g., 90%, representing good furnace visual conditions) and calculates the first deviation. Based on the magnitude and direction of the first deviation, and in accordance with the preset control strategy, control and adjustment commands for the corresponding operating variables (such as hydrogen flow rate, trichlorosilane flow rate, etc.) can be generated.

[0077] For example, some basic control methods include PI control (proportional-integral control) and PID control. Taking PI control as an example, the specific control strategy can be as follows: ,in It is the adjustment amount for the operated variable. and These are the controller parameters. Additionally, some more advanced control methods can combine model predictive control (MPC) to minimize the clarity bias of future multi-step predictions, and to continuously optimize the calculation of the optimal adjustment sequence of the manipulated variables.

[0078] The generated control commands can be sent to specific actuators (such as current regulators and flow control valves) through the control system to achieve closed-loop regulation of the clarity inside the furnace.

[0079] The polysilicon reduction furnace control method provided in this embodiment can provide real-time estimates of furnace clarity based on a high-precision dynamic soft-sensor model, thereby achieving closed-loop feedback control of this key quality indicator. This method solves the problem of traditional control systems being unable to perform precise regulation due to the lack of direct clarity measurement signals. It enables operators or automated systems to proactively intervene in the production process based on clear quantitative indicators, effectively suppressing furnace atomization and improving silicon rod deposition quality and production safety.

[0080] In one embodiment, after inputting the real-time operating parameters into the soft sensor model in step S204, the polysilicon reduction furnace control method further includes: obtaining the estimated surface temperature of the silicon rod output by the soft sensor model; at this time, the step of generating control instructions for the manipulated variable in step 206 may specifically include: generating control instructions for the manipulated variable based on a first deviation and a second deviation between the estimated surface temperature of the silicon rod and the temperature setpoint. This embodiment proposes a control instruction generation method that comprehensively considers the first and second deviations, thereby achieving multi-variable coordinated control of temperature and clarity. This method effectively avoids adverse interference to other variables when controlling one variable alone, and is key to improving overall process performance.

[0081] In one embodiment, generating control instructions for the operating variables based on the first deviation and the second deviation includes: if the absolute value of the second deviation exceeds a temperature threshold, generating a first control instruction for the heating current based on the estimated surface temperature of the silicon rod and a temperature setpoint; if the absolute value of the second deviation does not exceed the temperature threshold and the absolute value of the first deviation exceeds a clarity threshold, generating a second control instruction for the hydrogen flow rate and / or trichlorosilane flow rate based on the estimated clarity in the furnace and a clarity setpoint.

[0082] In one embodiment, the real-time operating parameters include the current silicon rod diameter. The polysilicon reduction furnace control method further includes: if the absolute value of the second deviation does not exceed the temperature threshold and the absolute value of the first deviation does not exceed the clarity threshold, then obtaining the third deviation between the current silicon rod diameter and the target silicon rod diameter; and generating a third control command for the trichlorosilane flow rate based on the third deviation.

[0083] The above embodiment refines the priority control logic based on temperature and sharpness deviation. Its core control principle is: 1) Safety First: When the temperature deviation exceeds a safety threshold (e.g., 5°C), indicating a risk of overheating and melting, an adjustment command for the heating current is generated first to quickly stabilize the temperature. Specifically, it can be implemented according to... Reduce current, Set the temperature to the above value, such as 1100℃.

[0084] 2) Quality Priority: When the temperature is normal but the clarity deviation exceeds the standard (e.g., exceeding 5%), indicating that atomization is severely affecting the quality, an adjustment command is generated to adjust the hydrogen flow rate and / or trichlorosilane flow rate. This aims to enhance purging or suppress particle formation and improve clarity. Specific details can be found by... Increase hydrogen flow rate, and simultaneously... Reducing the flow rate of trichlorosilane suppresses the formation of atomized particles; among which Set the resolution value as described above, such as 90%.

[0085] 3) Efficiency Priority: When temperature and clarity are well controlled, an adjustment command for the trichlorosilane flow rate is generated based on the deviation between the current silicon rod diameter and the target silicon rod diameter. This aims to optimize the deposition rate and improve production efficiency. Specifically, it can be configured according to... Increase the flow rate of trichlorosilane, among which The target silicon rod diameter is as described above.

[0086] This hierarchical prioritization strategy clarifies the primary control objectives and corresponding operational methods under different operating conditions, making the control strategy more secure, robust, and practical.

[0087] In one embodiment, prior to S204, the polysilicon reduction furnace control method further includes: determining the current growth stage of the polysilicon reduction furnace operation process based on the current silicon rod diameter; the step of inputting real-time operating parameters into the soft sensor model in S204 includes: inputting the real-time operating parameters into the soft sensor model corresponding to the current growth stage.

[0088] In this embodiment, before inputting real-time operating parameters into the soft measurement model, the current growth stage to which the silicon rod belongs can be determined based on the current silicon rod diameter, and then the dedicated model corresponding to that stage can be called for estimation. This scheme can ensure that the state estimation on which the controller depends is optimal for the current stage throughout the entire growth cycle, thereby ensuring the consistency of control accuracy from beginning to end and avoiding the decline in control performance caused by model mismatch.

[0089] In one embodiment, the polysilicon reduction furnace control method further includes: performing a first verification on a soft sensor model at a first frequency; the first verification is performed based on an in-furnace sharpness estimate, a calibration sharpness estimate, a silicon rod surface temperature estimate, and a calibration temperature, wherein the calibration sharpness and calibration temperature are obtained based on calibration data consistent with the current operating conditions in a historical database; performing a second verification on the soft sensor model at a second frequency; the second verification is performed based on an in-furnace sharpness estimate, a visual sharpness estimate, a silicon rod surface temperature estimate, and a visual temperature, wherein the visual sharpness and visual temperature are obtained by processing in-furnace images; wherein the first frequency is greater than the second frequency.

[0090] In a specific embodiment, performing a first verification includes: acquiring a first absolute deviation between the furnace sharpness estimate and the calibrated sharpness, and a second absolute deviation between the silicon rod surface temperature estimate and the calibrated temperature; if the first absolute deviation is greater than a first sharpness verification threshold, or the second absolute deviation is greater than a first temperature verification threshold, then the following steps are executed repeatedly until a first stopping condition is met: calling a preset number of recent historical running data to retrain the model structure to be trained, obtaining a new soft measurement model; based on the new soft measurement model, re-acquiring the corresponding furnace sharpness estimate and silicon rod surface temperature estimate, and calculating new first absolute deviations and second absolute deviations; wherein, the first stopping condition includes: the new first absolute deviation is not greater than the first sharpness verification threshold, and the new second absolute deviation is not greater than the first temperature verification threshold; and the number of times the loop is executed reaches a preset number.

[0091] In one specific embodiment, performing a second verification includes: acquiring visual clarity and visual temperature based on in-furnace images; acquiring a first relative deviation between the in-furnace clarity estimate and the visual clarity, and a second relative deviation between the silicon rod surface temperature estimate and the visual temperature; if the first relative deviation is greater than a second clarity verification threshold, or the second relative deviation is greater than a second temperature verification threshold, then continuously acquiring multiple new in-furnace images and correspondingly calculating multiple first relative deviations and second relative deviations; calculating a first average value of multiple sets of first relative deviations, and calculating a second average value of multiple sets of second relative deviations; if the first average value is greater than a second clarity verification threshold, or the second average value is greater than a second temperature verification threshold, then determining that the current soft measurement model is abnormal.

[0092] The above embodiments are intended to illustrate in detail the specific implementation process of the two-level verification mechanism, which is a safety redundancy design to ensure the long-term reliable operation of the control system.

[0093] Specifically, the first verification, namely model self-verification, is performed at a relatively high frequency (e.g., every minute). Its core is to compare the current output of the soft measurement model with the calibration data under the same operating conditions in the historical database. If the deviation exceeds the threshold, it triggers the adaptive update of the model parameters or an alarm. This verification can be used for online monitoring and maintaining the accuracy of the soft measurement model itself.

[0094] The second verification can be understood as a soft measurement-vision cross-verification, which is performed at a lower frequency (e.g., every 3 minutes). The core of this verification is to cross-compare the soft measurement output with the measurement results of an independent machine vision system. If the deviation between the two continues to exceed the limit, it is determined that the soft measurement model may have failed or the vision system may have been severely interfered with. This immediately triggers a safety switch, switching the controller from an advanced control mode that relies on soft measurement (e.g., MPC) to a traditional backup control system that does not rely on soft measurement (e.g., PID).

[0095] The aforementioned verification mechanism constructs a complete security defense, greatly improving the reliability and fault tolerance of the control system and effectively solving the problem of the lack of a model verification mechanism.

[0096] The following is a specific example for constructing a dynamic soft sensor model: In this example, the dynamic soft sensor model adopts a hybrid model of physical constraint model and model predictive control algorithm, with a 10-60s lag compensation term added, the ratio of training set to test set is 7:3, and the model evaluation index is MAE (mean absolute error) ≤1℃.

[0097] Furthermore, the variable mapping relationship is clarified using the MIMO vector function form: Soft measurement output vector: ,in The surface temperature of the silicon rod. For in-furnace sharpness, the unit of in-furnace sharpness is %; soft measurement input vector: , The heating current is in A; The hydrogen flow rate is m³ / h. The diameter of the silicon rod is in mm; Heating power, in kW; Cooling water flow rate, m³ / h; Lag compensation term: =10-30s: indicates the hysteresis of current to temperature; =20-40s: indicates the lag between the flow rate and the resolution; =30-60s: Indicates the hysteresis of the diameter on the coupling relationship, determined based on historical data fitting; Estimated atmosphere temperature: Where c is the specific heat capacity of water, 4.2 kJ / (kg· ... ℃); The temperature difference between the inlet and outlet of the cooling water, in °C; The furnace heat loss is expressed in kW; m is the mass of gas inside the furnace, expressed in kg. The specific heat capacity of the gas is 1.4 kJ / (kg). ℃); the final soft measurement vector function: ;in, , These are model coefficients, which can be obtained through training with 10,000+ sets of historical data, such as in Stage 1: , ; , For error terms (≤±2%, ≤±3%).

[0098] For atmosphere temperature estimation: the energy balance formula can be used: Q1 (heating power) = Q2 (heat removed by cooling water) + Q3 (heat absorbed by the atmosphere) + Q4 (heat loss). Q3 is used to calculate the atmosphere temperature, which is then incorporated into the furnace atomization detection model. For stage modeling and control implementation: stage 1 is defined as silicon rod diameter not exceeding 50mm, stage 2 as 50-100mm, and stage 3 as not less than 100mm. Independent models are established for each stage. After controller configuration, offline simulation is performed. During online implementation, the diameter data is updated every 3 minutes to achieve dynamic model adaptation. For control triggering conditions: control is triggered when any of the following conditions are met: temperature deviation: ( (Temperature setting, such as 1100℃); Sharpness deviation: ( (Set a resolution value, such as 90%); diameter variation: (The silicon rod diameter growth exceeds the threshold, and the stage model needs to be switched).

[0099] Regarding the determination of control target priority: Priority 1 (safety first): if (Overheating risk) Prioritize adjusting the heating current. ,according to Reduce current to prevent the silicon rod from melting; Priority 2 (quality first): If Normal but (Excessive atomization), adjust hydrogen flow rate. (according to Increase flow rate), and press Reduce the flow rate of trichlorosilane to suppress the formation of atomized particles; Priority 3 (efficiency first): if , All are normal, according to Increase the flow rate of trichlorosilane ( (to target diameter), to increase deposition rate.

[0100] Regarding control command output and execution: The model predictive controller calculates the adjustment amount of the manipulated variable based on the above logic, and outputs it to the actuator via D / A conversion: when current regulation is required, the command is sent to the current regulator with an adjustment accuracy of ±0.1A; when flow regulation is required, the command is sent to the flow regulating valve with an adjustment accuracy of ±0.01m³ / h; feedback data is collected within 100ms after execution to determine... Check if it reverts to the target range; if not, repeat the iterative adjustment.

[0101] For dual calibration and switching: when the deviation between soft measurement and visual inspection is greater than 5%, the predictive controller is disconnected and switched to the PID backup system to ensure production safety.

[0102] Specifically, for the self - calibration of the soft - sensor model (frequency: once every 1 minute), the source of calibration data is: retrieve the standard calibration data (offline calibrated by a high - precision infrared sensor) in the historical database that is consistent with the current working condition ( and and ) with an error ≤ 5%); the calibration indicators are: temperature deviation: ( is the calibrated temperature), and the qualified threshold ≤ 1°C; clarity deviation: ( is the calibrated clarity), and the qualified threshold ≤ 3%; the processing logic is: 1. If and , determine that the model is valid and maintain the control state; 2. If any deviation exceeds the standard, call the last 100 sets of valid data to refit the soft - sensor function coefficients (such as ), and re - calibrate after updating; 3. If the calibration fails continuously for 3 times, issue an abnormal warning for the model and prompt to check the sensor data.

[0103] Furthermore, for the double - calibration of soft - sensor - visual detection (frequency: once every 3 minutes), the method for obtaining visual detection data is: collect the images inside the furnace through a visual camera, and calculate as the visual temperature and the visual clarity through image segmentation (based on the gray - level threshold method) and particle - counting algorithm; the calibration indicators are: relative deviation , the qualified threshold for temperature ≤ 5%, and the qualified threshold for clarity ≤ 8%; the processing logic is: 1. If , determine that the double - calibration is qualified and maintain the model predictive control; 2. If any deviation exceeds the standard, continuously collect 3 sets of data (with an interval of 10 s) to calculate the average deviation : If is within the qualified threshold, it is determined as an instantaneous interference and the original control is maintained; if exceeds the qualified threshold, immediately trigger a switching instruction to cut off the model predictive controller and switch to the PID standby system; 3. PID system parameters: temperature PID (K p = 2.5, K i = 0.05, K d = 0.8), flow PID (K p = 1.2, K i = 0.03, K d = 0.3). During operation, re - double - calibrate every 30 s, and switch back to the model predictive control if 5 consecutive sets are qualified.

[0104] This embodiment fully demonstrates the closed-loop process from dynamic soft-sensor modeling and phased adaptation to multi-variable priority collaborative control and two-level safety verification. This method can stabilize the silicon rod surface temperature control accuracy within ±2℃ and control the furnace clarity estimation error within ±3%. Simultaneously, priority collaborative control effectively reduces multi-variable coupling interference; the two-level verification mechanism enables high-quality model anomaly monitoring and rapid safety controller switching in the event of model anomalies. This solution comprehensively improves the quality control level, operational efficiency, and intrinsic safety of polysilicon reduction production.

[0105] The deficiencies of the above solutions and the proposed solutions are the result of the inventor's practice and careful research. Therefore, the discovery process of the above problems and the solutions proposed in this disclosure below should be considered as the inventor's contribution to this disclosure.

[0106] It should be understood that, for the foregoing method embodiments, although the steps in the flowcharts are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the method embodiments may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps.

[0107] In one embodiment, a computer device is provided, the internal structure of which can be shown in Figure 3. The computer device includes a processor, memory, input / output interface, communication interface, display unit, and input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The input / output interface is used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, NFC (Near Field Communication), or other technologies. When the computer program is executed by the processor, it implements a polysilicon reduction furnace control method. The display unit is used to form a visually visible image and can be a display screen, projection device, or virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.

[0108] Those skilled in the art will understand that the structure shown in Figure 3 is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or may combine certain components, or may have different component arrangements.

[0109] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0110] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification. In the above embodiments, the descriptions of each embodiment have their own emphasis. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0111] The terms “comprising” and “having”, and any variations thereof, in the embodiments herein are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or (module) units is not limited to the steps or units listed, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such process, method, product, or apparatus.

[0112] In this article, "multiple" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. The character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0113] The terms "first" and "second" used herein are merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first" and "second" can be interchanged in a specific order or sequence where permissible. It should be understood that the objects distinguished by "first" and "second" can be interchanged where appropriate so that the embodiments described herein can be implemented in an order other than those illustrated or described herein.

[0114] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for constructing a soft-sensor model of a polycrystalline silicon reduction furnace, characterized in that, include: Determine the set of input variables and the set of output variables; the set of output variables includes furnace clarity, and the set of input variables includes hydrogen flow rate, silicon rod diameter, heating power, and cooling water flow rate; A set of lag time constants is determined to characterize at least the lag effect of the hydrogen flow rate and / or the silicon rod diameter on the furnace sharpness; a furnace atmosphere temperature estimation sub-model is constructed to calculate the furnace atmosphere temperature estimate based on the heating power, the cooling water flow rate, and the furnace body heat loss; a training model structure is constructed based on the input variable set, the output variable set, the lag time constant set, and the furnace atmosphere temperature estimation sub-model; the training model structure is trained based on historical operating data of the polysilicon reduction furnace to obtain the corresponding soft measurement model; the soft measurement model is configured to receive real-time data from the input variable set and output at least the furnace sharpness estimate.

2. The method according to claim 1, characterized in that, The set of output variables also includes the silicon rod surface temperature, and the set of input variables also includes the heating current; the set of hysteresis time constants is also used to characterize the hysteresis effect of the heating current on the silicon rod surface temperature; the soft measurement model is also configured to output an estimated value of the silicon rod surface temperature.

3. The method according to claim 2, characterized in that, The set of hysteresis time constants includes a first hysteresis time constant, a second hysteresis time constant, and a third hysteresis time constant; wherein, the first hysteresis time constant is used to characterize the hysteresis time of the effect of heating current change on silicon rod surface temperature; the second hysteresis time constant is used to characterize the hysteresis time of the effect of hydrogen flow rate change on furnace clarity; and the third hysteresis time constant is used to characterize the hysteresis time of the effect of silicon rod diameter change on model coupling relationship.

4. The method according to claim 3, characterized in that, The first lag time constant has a value range of 10-30 seconds, the second lag time constant has a value range of 20-40 seconds, and the third lag time constant has a value range of 30-60 seconds.

5. The method according to claim 3, characterized in that, The furnace atmosphere temperature estimation sub-model calculates the estimated furnace atmosphere temperature based on the following formula. ： ;in, The heating power; The cooling water flow rate; This refers to the specific heat capacity of water. The temperature difference between the inlet and outlet of the cooling water; This refers to the heat loss of the furnace body; The mass of gas inside the furnace; This represents the specific heat capacity of the gas.

6. The method according to claim 5, characterized in that, The structure of the model to be trained is as follows: ;in, The surface temperature of the silicon rod; For the clarity inside the furnace; The heating current; The hydrogen flow rate; The diameter of the silicon rod; These are the model coefficients; 、 This is the error term; The first lag time constant; This is the second lag time constant; The third lag time constant is mentioned above.

7. The method according to any one of claims 1 to 6, characterized in that, The method further includes: dividing the polysilicon reduction furnace operation process into multiple growth stages corresponding to the silicon rod diameter ranges according to multiple preset silicon rod diameter ranges; and independently training a corresponding soft measurement model for each growth stage.

8. The method according to claim 7, characterized in that, The multiple growth stages include a first stage, a second stage, and a third stage; the silicon rod diameter range corresponding to the first stage is no greater than 50 mm, the silicon rod diameter range corresponding to the second stage is greater than 50 mm and less than 100 mm, and the silicon rod diameter range corresponding to the third stage is no less than 100 mm.

9. A method for controlling a polycrystalline silicon reduction furnace, characterized in that, include: Obtain real-time operating parameters of the polysilicon reduction furnace; input the real-time operating parameters into the soft measurement model as described in any one of claims 1 to 8 to obtain at least the furnace sharpness estimate; generate control instructions for the operating variables based at least on a first deviation between the furnace sharpness estimate and the sharpness setpoint.

10. The method according to claim 9, characterized in that, After inputting the real-time operating parameters into the soft sensor model, the method further includes: obtaining the estimated surface temperature of the silicon rod output by the soft sensor model; the step of generating control instructions for the operating variable includes: generating control instructions for the operating variable based on the first deviation and the second deviation between the estimated surface temperature of the silicon rod and the temperature setpoint.

11. The method according to claim 10, characterized in that, Based on the first deviation and the second deviation, control instructions for the operating variables are generated, including: if the absolute value of the second deviation exceeds a temperature threshold, a first control instruction for the heating current is generated based on the estimated surface temperature of the silicon rod and the temperature setpoint; if the absolute value of the second deviation does not exceed the temperature threshold, and the absolute value of the first deviation exceeds a clarity threshold, a second control instruction for the hydrogen flow rate and / or trichlorosilane flow rate is generated based on the estimated clarity in the furnace and the clarity setpoint.

12. The method according to claim 11, characterized in that, The real-time operating parameters include the current silicon rod diameter. The method further includes: if the absolute value of the second deviation does not exceed the temperature threshold and the absolute value of the first deviation does not exceed the resolution threshold, then obtaining a third deviation between the current silicon rod diameter and the target silicon rod diameter; and generating a third control command for the trichlorosilane flow rate based on the third deviation.

13. The method according to claim 12, characterized in that, Before inputting the real-time operating parameters into the soft sensor model, the method further includes: determining the current growth stage of the polysilicon reduction furnace operation process based on the current silicon rod diameter; the step of inputting the real-time operating parameters into the soft sensor model includes: inputting the real-time operating parameters into the soft sensor model corresponding to the current growth stage.

14. The method according to any one of claims 10 to 13, characterized in that, The method further includes: performing a first verification on the soft measurement model at a first frequency; the first verification is performed based on the furnace sharpness estimate, calibration sharpness, silicon rod surface temperature estimate, and calibration temperature, wherein the calibration sharpness and calibration temperature are obtained based on calibration data consistent with the current operating conditions in a historical database; performing a second verification on the soft measurement model at a second frequency; the second verification is performed based on the furnace sharpness estimate, visual sharpness, silicon rod surface temperature estimate, and visual temperature, wherein the visual sharpness and visual temperature are obtained by processing furnace images; the first frequency is greater than the second frequency.

15. The method according to claim 14, characterized in that, Performing the first verification includes: obtaining a first absolute deviation between the furnace sharpness estimate and the calibrated sharpness, and a second absolute deviation between the silicon rod surface temperature estimate and the calibrated temperature; if the first absolute deviation is greater than a first sharpness verification threshold, or the second absolute deviation is greater than a first temperature verification threshold, then the following steps are executed repeatedly until a first stopping condition is met: calling a preset number of recent historical running data to retrain the model structure to be trained to obtain a new soft measurement model; based on the new soft measurement model, re-obtaining the corresponding furnace sharpness estimate and silicon rod surface temperature estimate, and calculating a new first absolute deviation and a new second absolute deviation; wherein, the first stopping condition includes: the new first absolute deviation is not greater than the first sharpness verification threshold, and the new second absolute deviation is not greater than the first temperature verification threshold; and the number of times the loop is executed reaches a preset number.

16. The method according to claim 14, characterized in that, Performing the second verification includes: acquiring the visual clarity and visual temperature based on the furnace image; acquiring a first relative deviation between the furnace clarity estimate and the visual clarity, and a second relative deviation between the silicon rod surface temperature estimate and the visual temperature; if the first relative deviation is greater than a second clarity verification threshold, or the second relative deviation is greater than a second temperature verification threshold, then continuously acquiring multiple new furnace images and calculating multiple first relative deviations and second relative deviations accordingly; calculating a first average value of the multiple sets of first relative deviations, and calculating a second average value of the multiple sets of second relative deviations; if the first average value is greater than the second clarity verification threshold, or the second average value is greater than the second temperature verification threshold, then determining that the current soft measurement model is abnormal.

17. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 8 or claims 9 to 16.

Images