Multi-parameter nonlinear prediction method for antimony vacuum casting process

Abstract

The invention relates to the technical field of metal vacuum casting, and provides a multi-parameter nonlinear prediction method for an antimony vacuum casting process, which comprises the following steps of: converting a predicted crystal interface and impurity distribution result into a control objective function with a space partition weight and a risk punishment mechanism; and the control strategy can be differentially optimized according to the quality sensitivities of different areas. On this basis, a trainable control strategy generation mechanism is designed, so that the control strategy generation mechanism outputs heating power, cooling flux and crystal ingot pulling speed adjustment instructions which can be directly executed by minimizing future prediction target deviation, and confidence perception amplitude limiting, state sensitive correction and strategy backspacing mechanisms are introduced during deployment. And it is ensured that the control system still has safety and stability in a complex disturbance environment. According to the method, systematic structural innovation is realized in three links of prediction, target construction and execution, and the core problem that prediction and control are separated and closed-loop operation cannot be realized in a traditional method is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of metal vacuum casting, in particular to a multi-parameter nonlinear prediction method for antimony vacuum casting process. BACKGROUND

[0002] As an important functional material, the vacuum casting process of antimony directly affects the crystallization interface morphology and internal impurity distribution of the ingot, and has a decisive role on the final crystal purity and product quality.

[0003] In the prior art, process control is mostly dependent on temperature curves set by experience or simplified models based on local parameters, and in actual production, it often faces two problems, one is the lack of prediction ability, which cannot accurately represent the evolution trajectory of the crystallization interface and the migration trend of impurities in the crystal, especially under complex working conditions, which is prone to prediction distortion, the other is the lag of control strategy, which usually relies on manual adjustment or fixed proportional regulation, lacks organic connection with the prediction model, resulting in insufficient control accuracy, easy to appear interface disturbance and impurity aggregation phenomenon, affecting the yield and purity of the ingot. Some studies try to improve the prediction effect by coupling numerical simulation of heat conduction and material migration, but the calculation cost is high, the real-time performance is poor, and the results are difficult to directly guide the temperature control strategy generation. On the other hand, although the machine learning method based on data driving alone has fitting ability, it lacks physical constraints, and the prediction results often do not match the actual process, which is difficult to obtain industrial recognition.

[0004] Therefore, the existing prediction and control system still has problems such as low precision, insufficient reliability, unstable response, etc. in the antimony vacuum casting process, and an accurate prediction method combining physical constraints and data driving is needed, and on this basis, a control strategy generation and execution mechanism is formed, so as to realize accurate control of the crystal growth process. SUMMARY

[0005] The present application provides a multi-parameter nonlinear prediction method for antimony vacuum casting process, in order to solve the problems of inaccurate prediction and control lag in the prior art, a nonlinear prediction model is constructed with process multi-parameter data as input, and physical constraints such as thermal field gradient, pull speed influence and impurity migration under vacuum condition are explicitly introduced into the model structure, so as to obtain a prediction result balanced between physical rationality and data fitting ability.

[0006] To achieve the above purpose, the technical scheme adopted by the present application is: A multi-parameter nonlinear prediction method for antimony vacuum casting process is provided, comprising the following steps: Collect key process state variables of antimony in the vacuum melting process and preprocess to obtain a pretreated single-time process state vector, and introduce a historical window mechanism to construct a time series tensor to obtain a historical process state sequence tensor at the past time, the key process state variables including temperature distribution, vacuum degree, crystal ingot moving speed and current position of the crystal ingot, the preprocessing steps including normalization and structure alignment operations; Construct a nonlinear prediction model based on the historical process state sequence tensor, the nonlinear prediction model including two channels, two channels respectively processing temperature sequence and position and speed vacuum channels, two channels respectively used for outputting a predicted future time crystallization interface position vector and an impurity distribution vector; Based on the future time crystallization interface position vector and the impurity distribution vector, a target function for control strategy optimization is constructed, and based on the target function, a control strategy function is constructed, the control strategy function used for outputting adjustment instructions by learning the functional relationship between the current process state and the predicted target; Based on the control strategy function and the adjustment instructions, the final control instructions are converted through a dynamic mapping function, the final control instructions are issued to the equipment end to realize intelligent control of the heater, the cooling unit and the crystal ingot traction system, the final control instructions including heating power module control amount, cooling water flow control amount and crystal ingot pulling speed control amount.

[0007] As preferred, the temperature distribution is obtained by collecting a thermocouple array installed in the melting cavity; the vacuum degree is obtained by collecting an ionization vacuum gauge at the top of the furnace body; the crystal ingot moving speed is obtained by real-time output of an encoder in the servo system; and the current position of the crystal ingot is obtained by direct measurement of a laser range finder or a displacement meter.

[0008] As preferred, the preprocessing operation steps are as follows: All key process state variables are aligned to a set uniform sampling period, for data with a sampling frequency higher than the period, a moving average is used for downsampling; and for data lower than the period, linear interpolation is used for padding, and time synchronization is unified by controller master clock for unified timestamp binding.

[0009] As preferred, the nonlinear prediction model introduces a loss term to ensure that the predicted crystallization interface advancing speed is consistent with the current thermal field gradient and the crystal ingot pulling speed in physical sense and to suppress non-physical changes in the impurity concentration prediction results.

[0010] As preferred, the control strategy function adopts a three-layer fully connected network structure, the first layer has 64 nodes, the second layer has 32 nodes, and the third layer output layer has 3 nodes corresponding to control variables, all activation functions are LeakyReLU, and the output is a normalized control amount.

[0011] More preferably, in the control strategy function training process, the predicted future target function value is taken as the optimization object, the control strategy function is trained, the control strategy function parameters are frozen after the training, and the control strategy function parameters are deployed in the vacuum melting and casting control system, the control adjustment amount is output in real time according to the current process state input every second, and the prediction model is cyclically run.

[0012] Preferably, the adjustment instruction includes adjustment increments of heating power, cooling rate and crystal ingot moving speed.

[0013] Preferably, the heating power module control amount is used for input to a heater power module to control the output current of a graphite heating element, the cooling water flow control amount is used for input to a cooling water flow controller to control the cooling water flux, and the crystal ingot pulling speed control amount is used for input to a crystal ingot traction servo system to control the crystal ingot pulling speed.

[0014] More preferably, the heating power module control amount, the cooling water flow control amount and the crystal ingot pulling speed control amount are written into an industrial control system through an analog quantity control interface to directly drive corresponding execution devices.

[0015] Preferably, before the final control instruction is executed, an abnormal prediction offset inhibition term is introduced to modify the adjustment instruction to obtain the final control instruction; when the prediction target function is extremely sensitive to the input state, it indicates that the prediction behavior is unstable, and at this time, the amplitude of the control instruction should be automatically reduced.

[0016] Compared with the prior art, the present application has the following beneficial effects: The present application proposes a prediction and control integrated scheme for the antimony vacuum melting and casting process, constructs a nonlinear prediction model with process multi-parameter data as input to solve the problems of inaccurate prediction and control lag, and explicitly introduces physical constraints such as thermal field gradient, pulling speed influence and impurity migration under vacuum condition in the model structure to obtain a prediction result balanced between physical rationality and data fitting capability. Further, the present application converts the predicted crystallization interface and impurity distribution result into a control target function with spatial partition weight and risk penalty mechanism, so that the control strategy can be optimized differently according to the quality sensitivity of different regions. On this basis, a trainable control strategy generation mechanism is designed to output directly executable heating power, cooling flux and crystal ingot pulling speed adjustment instructions by minimizing the future prediction target deviation, and to introduce confidence awareness limiting, state sensitive modification and strategy rollback mechanism when deployed to ensure the safety and stability of the control system under complex disturbance environment. The present application realizes systematic structural innovation in the prediction, target construction and execution three links, solves the core problem that the traditional method cannot be closed-loop run due to the separation of prediction and control, and can significantly improve the crystallization interface stability of the crystal ingot and the control precision of the antimony purity, thereby providing reliable process guarantee for high-quality crystal preparation. Attached Figure Description

[0017] Figure 1 This is a flowchart of a multi-parameter nonlinear prediction method for an antimony vacuum casting process in a specific embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] This invention provides a multi-parameter nonlinear prediction method for the antimony vacuum casting process, the method comprising: S1. Collect key process state variables of antimony in vacuum melting and casting process and preprocess them to obtain the preprocessed single-moment process state vector. Introduce the historical window mechanism to construct the time series tensor to obtain the historical process state sequence tensor of past moments. The key process state variables include temperature distribution, vacuum degree, ingot moving speed and current position of ingot. The preprocessing steps include normalization and structure alignment operations. This step aims to construct a standardized input data structure for predicting crystal behavior. This structure should accurately represent key process state variables during the antimony vacuum casting process, including temperature distribution, vacuum level, ingot movement speed, and current ingot position. To ensure the physical consistency and temporal uniformity of the model input, data with different sampling frequencies and physical magnitudes need to be normalized and structurally aligned so that they can be uniformly encapsulated into tensors suitable for sequence modeling.

[0020] Temperature distribution It is collected by a thermocouple array installed inside the casting cavity, usually along the ingot axis. Multiple temperature measurement points, such as 8 to 16, are evenly spaced along the coordinate axis. Data sampling frequency is approximately once every 0.5–2 seconds. This variable is the primary controlling factor for subsequent prediction of crystal interface morphology changes. Vacuum degree. Provided by an ionization vacuum gauge at the top of the furnace, the sampling frequency is typically once per second, reflecting real-time changes in furnace chamber pressure and significantly influencing impurity migration trends. Ingot movement speed. The output is provided in real time by the encoder in the servo system, with a high update frequency, typically once every 10–50 milliseconds. Current position of the ingot. Can be derived from The integral can be obtained, or it can be directly measured using a laser rangefinder or displacement meter.

[0021] In order to ensure data consistency, all original signals will be uniformly aligned to the set uniform sampling period , which can be set to 1 second. For data with a sampling frequency higher than this period, a sliding average is used for downsampling; for data with a lower sampling frequency, linear interpolation is used to fill in the gaps. Time synchronization is achieved through the controller's master clock, which binds the timestamps uniformly and avoids information drift caused by multiple asynchronous sources.

[0022] On this basis, the normalized input tensor is constructed as follows: ; , where is the pre-processed single-time process state vector, which serves as the representation of the current control state; represents the temperature value at the th temperature measurement point along the axial direction of the crystal ingot, which has been normalized to ensure dimensional consistency; is the normalized vacuum degree; is the normalized crystal ingot pulling speed; is the normalized current position. All variable normalization operations use a linear normalization function based on the maximum and minimum values of historical data to ensure uniformity of the model input feature scale.

[0023] In order to construct a time series modeling structure, a historical window mechanism is introduced to construct a time series tensor: ; is the historical process state sequence tensor of the past seconds, which represents the sequence stacking of continuous normalized process state samples from time to the current time , and is the standard input form of the subsequent nonlinear prediction model. Typically, , when the value is 20, it means that the model can perceive the continuous process evolution of the past 20 seconds.

[0024] For example, if the thermocouple temperature measurement points are set to , each contains temperature points, plus the vacuum degree, pulling speed, and position, totaling variables, and is set to , then the final tensor is a two-dimensional structure of .

[0025] The core innovation of the scheme is to combine the specific physical control logic of the antimony vacuum melting process, select only the key variables that affect the evolution of the crystallization behavior, and through a unified data structure construction method, all input data have clear spatial distribution and time evolution information, and are strictly aligned with the subsequent model structure. In particular, the windowed sequence modeling tensor can significantly enhance the model's ability to perceive the evolution trend of the crystal, improve the prediction accuracy, and reduce the fitting instability problems caused by data frequency or physical dimension differences.

[0026] S2, constructing a nonlinear prediction model based on the historical process state sequence tensor, the nonlinear prediction model including two channels, two channels respectively processing temperature sequence and position and speed vacuum channels, and the two channels being used for outputting a predicted future time crystallization interface position vector and an impurity distribution vector; The core task of this step is to construct a physically constrained nonlinear prediction model based on the historical process state sequence tensor output in step one, to predict the position distribution of the crystallization interface of the ingot at the future time and the axial concentration trend of the impurities Unlike general modeling methods, the present application fully considers the physical coupling characteristics, thermal field disturbance sensitivity, and discontinuity risk of impurity migration of the antimony vacuum melting process in the model structure and objective function design, and constructs a highly integrated prediction mechanism combining data-driven and engineering knowledge, to provide accurate and controllable prediction results for the generation of subsequent control strategies.

[0027] is a historical sequence tensor, with a dimension of , wherein the process state of each time step contains 10 axial temperature points and vacuum degree , ingot pulling speed and position three normalized scalars. The tensor fully reflects the thermal field distribution, vacuum environment and motion trajectory in the crystal growth process in the past seconds, which is a necessary information basis for realizing nonlinear dynamic modeling.

[0028] To fully utilize the time sequence structure, the model adopts a double-channel GRU (Gated Recurrent Unit) network structure as the backbone, and two channels process the temperature sequence and the position-speed-vacuum channel respectively, to separate high-frequency disturbances (such as temperature field fluctuations) and low-frequency process changes (such as ingot advancement). Each channel contains one layer of GRU, with 64 hidden units, and the final output is sent to the shared feature fusion layer after splicing, forming a dynamic encoding of the current system state ​On this basis, the model is split into two output branches, predicting the future h seconds after the crystallization interface position vector and impurity distribution vector . Each output branch is composed of two layers of fully connected networks, with ReLU activation, and the output length is , consistent with the temperature sampling points.

[0029] In the model training process, two physical constraint mechanisms specially designed for the scene of the present application are introduced, so that the prediction not only has data fitting ability, but also has physical interpretability and behavior consistency. The first innovative loss term is used to ensure that the predicted crystallization interface advancing speed is consistent with the current thermal field gradient and the crystal pulling speed in the physical sense: ; wherein is the current interface real position (training label), is the predicted value; is the normalized pulling speed; is the local temperature gradient approximation along the axial direction (using first-order difference); and are the adjustment coefficients of thermal field-motion coupling. This term reflects the thermal-kinetic double coupling relationship of the crystallization interface evolution, which is obvious in antimony casting, especially in the acceleration-deceleration crystal pulling stage.

[0030] The second loss term is used to suppress the non-physical sharp changes in the predicted results of impurity concentration. In the process of antimony vacuum casting, impurities often slowly move along the crystallization interface in the vacuum condition by floating + segregation mechanism, and the concentration gradient in the axial direction is usually low-speed continuous. Therefore, the following regularization term is introduced to limit the axial change rate of impurity concentration: ; wherein is the vacuum influence factor, is the normalized vacuum degree. This regularization term relaxes the restriction under high vacuum conditions, and increases the penalty under low vacuum or fluctuating conditions, reflecting the physical constraint that the prediction smoothness needs to be enhanced in the environment where impurities are more likely to be retained.

[0031] ; wherein is the mean square error between the two predicted quantities (interface position and impurity concentration) and the actual label, is the weight of the two regularization terms, which is obtained by cross-validation in the model validation stage.

[0032] The prediction of this step obtains the future seconds after the crystallization interface position vector and impurity distribution vector Both are of length. The vector will be used directly as the input for constructing the objective function and generating the control strategy in the next step.

[0033] S3. Construct an objective function for control strategy optimization based on the crystallization interface position vector and impurity distribution vector at future time. Construct a control strategy function based on the objective function. The control strategy function is used to output adjustment instructions by learning the functional relationship between the current process state and the predicted objective. This step aims to build upon the predicted crystallization interface evolution results of the ingot from the previous step. and impurity concentration distribution trend We construct an objective function that can be used for control strategy optimization and design an adaptive control strategy generation mechanism to achieve real-time adjustment of the temperature control system during the antimony vacuum casting process.

[0034] Unlike traditional temperature control systems that rely solely on temperature setpoints or simple deviations, this invention, for the first time, uses "predicted crystallization behavior" as the control target and expresses this target in a structured manner in axial space, enabling the control strategy to be specifically optimized for the behavior of different regions of the ingot.

[0035] This step is the key bridge for this invention to move from "nonlinear modeling" to "control execution". It constructs a functional connection structure between prediction and control, and drives the generation of control decisions through crystal quality indicators. It is the core link to realize the closed-loop control capability of the system.

[0036] To obtain the future prediction obtained in step S2 Crystallization interface position vector after seconds and impurity distribution vector Transforming the objective into an optimizable control objective, the following objective function is constructed: ; in, and This is the predicted value for the current location. For reference crystallization interfaces (which can be obtained from high-quality ingot slice images or numerical simulations), Target purity concentration. Weighting coefficient. and This indicates the priority of crystal morphology and impurity control at each spatial point. The setting can be adjusted according to the ingot's partitions; for example, impurities tend to accumulate at the head and tail of the ingot. Set higher values ​​in these areas. For innovative regional risk items, Representing a spatial point Risk-sensitive weights The "uncertainty of predicted behavior" in this region is obtained by estimating the variance of the hidden states in the prediction model, as follows: ; The introduction of this regularization term addresses the risk of localized temperature instability during the antimony vacuum casting process. Certain hot zones may experience reduced prediction reliability due to structural asymmetry or non-uniform heat transfer. By penalizing regions with high uncertainty, the control strategy can avoid over-reliance on locations with drastic fluctuations in predicted values, thereby improving the stability and safety of the entire control system. This regularization design differs from traditional L1 / L2 norms or KL divergence; it is a structured expression of model reliability based on industrial physical scenarios, reflecting the physics-model fusion characteristics of the control structure in this invention.

[0037] Based on the above objective function, construct the control strategy function. Used to determine the current process status Output the next adjustment command. The set of control variables is as follows: ; These three represent the adjustment increments of heating power, cooling rate, and ingot movement speed, respectively. The control strategy function adopts a three-layer fully connected network structure: the first layer has 64 nodes, the second layer has 32 nodes, and the third output layer has 3 nodes, corresponding to the control variables. All activation functions are LeakyReLU, and the output is a normalized control quantity.

[0038] During the training of the policy function, the current bias is not directly minimized; instead, the predicted future objective function value is used. To optimize the target, the following training objective is set: ; This expression indicates that the policy network learns the current process state. The functional relationship between the target and the prediction result is used to output an adjustment command that minimizes future crystal deviations, demonstrating forward-looking adjustment capabilities. This training is conducted in a supervised manner, with data sourced from "high-quality control paths" obtained through reverse engineering from historical high-quality batches, or simulated using temperature control curves set by process experts, ensuring that the strategy training samples are physically reasonable and structurally realistic.

[0039] Once the control strategy training is complete, it will be frozen. The parameters are deployed in the control system, and the control adjustment is output in real time every second based on the current process status. It works in conjunction with the predictive model to complete the closed loop of "prediction-evaluation-control".

[0040] This step yields two results: first, the value of the objective function is controlled. The first is used to characterize the degree of deviation between future crystal behavior and quality standards; the second is the control strategy function. Output adjustment instructions This serves as the input for the next control and execution module, driving the actual equipment to complete the process adjustment.

[0041] S4. Based on the control strategy function and adjustment instructions, the final control instructions are transformed into dynamic mapping functions and sent to the equipment to realize intelligent control of the heater, cooling unit and crystal traction system. The final control instructions include heating power module control quantity, cooling water flow control quantity and crystal pulling speed control quantity. This step involves using the control policy function obtained in the previous training step. and control objective function value Building upon this foundation, the key step in the system's transition "from modeling to control implementation" is to complete the industrial deployment and practical execution of the temperature control strategy. Unlike traditional melting and casting temperature control systems that operate solely based on a set temperature curve, this step enables the control system to automatically calculate adjustment commands based on predicted future crystal states and distribute them to the equipment via an industrially adaptable architecture, thereby achieving intelligent control of the heater, cooling unit, and ingot traction system.

[0042] To transform the policy output into actual executable instructions, a set of dynamic mapping functions with soft constraint mechanisms is first defined: ; in Indicates the first The final control command at that moment. This is the control value that was previously issued; For adjustment instructions; The unit transformation matrix maps the normalized control quantity to the actual adjustment range; Limit the rate of regulation to prevent sudden changes; To control the target deviation value, This represents the maximum historical deviation. To prevent the use of tiny constants with a denominator of zero, the above formula embodies an innovative mechanism: the better the prediction quality (i.e., the better the prediction quality), the more likely it is to result in a zero denominator. The smaller the prediction deviation, the more proactive the control behavior; when the prediction deviation is large, the control behavior tends to be conservative, avoiding overly aggressive control based on erroneous prediction values, reflecting the core idea of ​​the "confidence-weighted control strategy".

[0043] During the equipment execution phase, It will be broken down into three sub-control variables: : Heating power module control quantity, used to input to the heater power module to control the output current of the graphite heating element; Cooling water flow control quantity, used to input to the cooling water flow controller to control the cooling water flow rate; : Ingot pull-down speed control quantity, used to input to the ingot traction servo system to control the ingot pull-down speed.

[0044] The three control quantities mentioned above can be written into the industrial control system (such as PLC, DCS or embedded controller) through digital / analog control interfaces (such as 0-10V analog, RS485 or Profinet protocol) to directly drive the corresponding actuators.

[0045] To ensure industrial security and long-term operational stability of the deployment, this step proposes an anomaly prediction offset suppression term, used to suppress anomaly predictions before policy execution. Corrections are made to prevent equipment oscillations caused by short-term prediction errors. The correction items are defined as follows: ; in, This indicates that the control objective function is related to the current state variable. The gradient rate of change can be approximated by the Jacobian matrix during the policy training phase; This is an adjustable coefficient, representing the ability to tolerate prediction risks. This mechanism is essentially a state-sensitive control offset correction: when the prediction objective function is extremely sensitive to the input state, it indicates that the prediction behavior is unstable, and the amplitude of the control command should be automatically reduced.

[0046] In addition, to enhance long-term adaptability and fault recovery capabilities after deployment, a "policy rollback mechanism" is designed: when continuous Control objective function at each time step Higher than the set threshold The system will automatically pause the strategy control and switch to the temperature control curve preset by the expert. The current process is marked as an "abnormal control segment" for subsequent process backtracking and strategy fine-tuning optimization. This mechanism ensures that the casting process remains basically stable even under adverse conditions such as strategy overfitting and sudden external disturbances, reflecting the robustness considerations of this invention at the industrial deployment level. The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the invention. Various modifications and improvements to the technical solutions of the present invention made by those skilled in the art without departing from the spirit of the invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A multi-parameter nonlinear prediction method for antimony vacuum casting process, characterized in that, The method comprises the following steps: Collecting key process state variables of antimony during a vacuum melting process and preprocessing to obtain a preprocessed single-time process state vector, and introducing a historical window mechanism to construct a time series tensor to obtain a historical process state sequence tensor of past time, the key process state variables including temperature distribution, vacuum degree, crystal ingot moving speed and current crystal ingot position, and the preprocessing steps including normalization and structure alignment operations; Based on the historical process state sequence tensor, a nonlinear prediction model is constructed, the nonlinear prediction model including two channels, two channels respectively processing temperature sequence and position and speed vacuum channels, and two channels respectively used for outputting a predicted future time crystallization interface position vector and an impurity distribution vector; Based on the future time crystallization interface position vector and the impurity distribution vector, a target function for control strategy optimization is constructed, and based on the target function, a control strategy function is constructed, the control strategy function used for outputting adjustment instructions by learning the functional relationship between the current process state and the predicted target; Based on the control strategy function and the adjustment instructions, a dynamic mapping function is used to convert final control instructions, the final control instructions issued to the equipment end to realize intelligent control of the heater, the cooling unit and the crystal ingot traction system, the final control instructions including a heating power module control amount, a cooling water flow control amount and a crystal ingot pulling speed control amount.

2. A multi-parameter nonlinear prediction method for antimony vacuum casting process according to claim 1, characterized in that, The temperature distribution is obtained by collecting a thermocouple array installed in the melting cavity; the vacuum degree is obtained by collecting an ionization vacuum gauge at the top of the furnace body; the crystal ingot moving speed is obtained by real-time output of an encoder in the servo system; and the current crystal ingot position is obtained by direct measurement of a laser range finder or a displacement meter.

3. A multi-parameter nonlinear prediction method for antimony vacuum casting process according to claim 1, characterized in that, The preprocessing operation steps are as follows: All key process state variables are aligned to a set uniform sampling period, for data with a sampling frequency higher than the period, a moving average is used for downsampling; and for data lower than the period, linear interpolation is used for padding, and time synchronization is unified by controller master clock timestamp binding.

4. The multi-parameter nonlinear prediction method of antimony vacuum casting process according to claim 1, characterized in that, The nonlinear prediction model introduces a loss term to ensure that the predicted crystallization interface advancing speed is consistent with the current thermal field gradient and the crystal ingot pulling speed in the physical sense and to suppress non-physical changes in the impurity concentration prediction results.

5. A multi-parameter nonlinear prediction method for antimony vacuum casting process according to claim 1, characterized in that, The control strategy function adopts a three-layer fully connected network structure, the first layer has 64 nodes, the second layer has 32 nodes, and the third layer output layer has 3 nodes corresponding to control variables, all activation functions are LeakyReLU, and the output is a normalized control amount.

6. A multi-parameter nonlinear prediction method of antimony vacuum casting process according to claim 5, characterized in that, During the control strategy function training process, the predicted future target function value is taken as the optimization object, after the control strategy function training is completed, the control strategy function parameters are frozen and deployed in the vacuum melting control system, control adjustment amount is output in real time according to the current process state input every second, and the prediction model is circulated.

7. A multi-parameter nonlinear prediction method of antimony vacuum casting process according to claim 1, characterized in that, The adjustment instructions include adjustment increments of heating power, cooling rate and crystal ingot moving speed.

8. A multi-parameter nonlinear prediction method of antimony vacuum casting process according to claim 1, characterized in that, The heating power module control quantity is used for inputting to a heater power module to control the output current of the graphite heating element; the cooling water flow control quantity is used for inputting to a cooling water flow controller to control the cooling water flux; and the crystal ingot drawing speed control quantity is used for inputting to a crystal ingot traction servo system to control the crystal ingot drawing speed.

9. A multi-parameter nonlinear prediction method of antimony vacuum casting process according to claim 8, characterized in that, The heating power module control quantity, the cooling water flow control quantity and the crystal ingot drawing speed control quantity are written into an industrial control system through an analog quantity control interface to directly drive corresponding execution devices.

10. The multi-parameter nonlinear prediction method of antimony vacuum casting process according to claim 1, characterized in that, Before the final control instruction is executed, an abnormal prediction offset inhibition term is introduced to modify the adjustment instruction to obtain the final control instruction; when the prediction target function is extremely sensitive to the input state, it indicates that the prediction behavior is unstable, and at this time, the amplitude of the control instruction should be automatically reduced.

Images