Vacuum self-consumption process prediction control system

Abstract

The invention relates to a vacuum self-consuming process prediction control system, which is matched with the arrangement of a vacuum self-consuming furnace, collects original process data from the vacuum self-consuming furnace through a multi-sensor detection unit, receives the original process data through a state estimation and feature enhancement module unit and converts the original process data into enhanced feature data; and the prediction control unit executes online prediction based on the enhanced feature data, and finally controls the execution unit to output and execute power control and motion control. According to the invention, the accuracy and adaptability of model prediction are obviously improved; good universality and generalization are achieved, and a new technical thought is provided for improving the automation and intelligence level of the industrial smelting process; under the condition that the control requirements of the target melting speed and the inter-electrode distance are met, the fluctuation amplitude of the output current is reduced, and the system is more stable; and the real-time performance, the prediction precision and the robustness of the process state are obviously improved.

Description

Technical Field

[0001] This invention relates to the technical field of automation in special alloy smelting, and in particular to a predictive control system for vacuum self-consumption processes. Background Technology

[0002] Vacuum arc remelting (VAR) uses a metal or alloy ingot as a consumable electrode, suspended vertically above a vacuum furnace, with a water-cooled copper crystallizer below. After vacuuming, an electric arc is ignited. The arc heat melts the lower end of the consumable electrode, forming a droplet. The droplet falls into the crystallizer, rapidly cools, and solidifies into an ingot. The electrode slowly descends as it is consumed to maintain arc stability. Due to its flexible melting method and excellent material modification effects, vacuum AMR plays an irreplaceable role in the preparation of high-end materials.

[0003] Vacuum arc remelting (VAR) furnaces are now computer-controlled and can perform melting according to a predetermined program. In industrial production, the core control objective of the VAR process is to stabilize the melting rate and electrode spacing, both of which cannot be directly measured online. In existing technologies, these can only be inferred from easily measurable indirect variables (such as electrode quality, arc voltage, and droplet rate), but this correlation is weak and unstable. The instability of the VAR process arc results in voltage and current signals containing a large amount of random noise and fluctuating drastically. Furthermore, the VAR process is a complex process involving thermodynamics and fluid dynamics, exhibiting high nonlinearity, making it difficult for traditional linear models (such as PID controllers) to adapt. Therefore, accurately estimating these key internal states that cannot be directly measured is one of the urgent problems to be solved.

[0004] Secondly, it is also necessary to address how to extract stable and reliable features from noisy and nonlinear sensor data to guide control. When electrode specifications, alloy grades, or equipment conditions change, existing control methods based on simplified physical models often require recalibration of model parameters, resulting in poor generalization ability.

[0005] Finally, the vacuum arc remelting (VAR) process typically exhibits significant time delay characteristics, with different time delays for different state variables, such as melting rate lag and dripping rate lag. Due to these time delays, different control systems have varying analytical capabilities regarding the "state" of the arc remelting process. Currently used PID control systems can typically only acquire the "past" state of the arc remelting furnace and perform calculations, feedback, and control based on this "past" data. Current VAR control technology suffers from drawbacks including over-reliance on human experience, poor response to the complex nonlinear dynamics of the melting process, inability to precisely control the melting rate and inter-electrode spacing, and instability in the final ingot quality (such as compositional segregation, black spots, and white spots). Therefore, a predictive control model with strong adaptability and generalization capabilities is needed.

[0006] With the rapid development of deep learning technology, recurrent neural networks have become an important method for solving complex cognitive problems. For the vacuum arc remelting process, how to extract effective features and train a high-precision prediction model is one of the core issues currently being addressed in research. Summary of the Invention

[0007] This invention addresses the problems existing in the prior art and provides a predictive control system for the vacuum self-consuming process. Based on a volumetric Kalman filter algorithm, this control system estimates the "current" state of the self-consuming furnace through multi-sensor fusion and electrode melting kinetics calculation, and performs calculations, feedback, and control based on the estimated value. By introducing a recurrent neural network, an algorithm is designed to predict the future state of the self-consuming furnace, enabling the control system to respond in advance and ultimately achieve more powerful and flexible control.

[0008] The technical solution adopted in this invention is a vacuum self-consumption process prediction and control system, which is set up in conjunction with a vacuum self-consumption furnace. The system includes:

[0009] A multi-sensor detection unit is used to collect raw process data from the self-vacuum self-consuming furnace;

[0010] A state estimation and feature enhancement module unit is used to receive the original process data and convert it into enhanced feature data;

[0011] A prediction control unit performs online prediction based on the enhanced feature data;

[0012] A control and execution unit is used to output and execute power control and motion control.

[0013] Preferably, the multi-sensor detection unit includes a current sensor for acquiring the actual melting current value of the electrode, a voltage sensor for acquiring the arc voltage, a mass sensor for acquiring the electrode mass, and a position sensor for acquiring the electrode position; the electrode feed speed is obtained based on the change in electrode position within two consecutive sampling periods.

[0014] Preferably, the state estimation and feature enhancement module unit uses the physical model of the electrode melting process as the benchmark model, integrates the volumetric Kalman filter algorithm to filter the raw process data collected by the multi-sensor detection unit, and generates new internal state features.

[0015] The enhanced feature vector is obtained by fusing the output of the state estimation and feature enhancement module units.

[0016] Preferably, the reference model is a one-dimensional heat conduction model with moving boundaries. By combining the basic equations and boundary conditions of the alloy as an electrode, the relationship between the electrode melting rate and the thermal boundary layer change rate is obtained.

[0017] Preferably, the electrode mass change rate and electrode gap change rate are obtained based on the electrode melting rate.

[0018] Preferably, the internal state characteristics include electrode thermal boundary layer thickness, efficiency ratio coefficient, current bias, electrode feed rate bias, filling ratio, voltage bias, and melting rate.

[0019] Preferably, the predictive control unit predicts the optimal melting current value and electrode feed rate value for the next control cycle.

[0020] Preferably, the prediction control unit is a recurrent neural network.

[0021] Preferably, the control execution unit includes a power controller and a motion controller. The power controller controls the optimal melting current value to act between the electrode and the molten pool, and the motion controller controls the motor to act on the electrode so that its movement speed meets the electrode feed speed value.

[0022] This invention relates to a vacuum self-consuming process prediction and control system, which is set up in conjunction with a vacuum self-consuming furnace. The system uses a multi-sensor detection unit to collect raw process data from the vacuum self-consuming furnace, a state estimation and feature enhancement module unit to receive the raw process data and convert it into enhanced feature data, a prediction and control unit to perform online prediction based on the enhanced feature data, and finally a control execution unit to output and execute power control and motion control.

[0023] The beneficial effects of this invention are as follows:

[0024] (1) Reduce the noise and its impact on the observation vectors obtained by the sensor, extract and construct more accurate and richer features for the improved deep learning model to learn and train, and significantly improve the accuracy and adaptability of the model prediction;

[0025] (2) A recurrent neural network is introduced to solve the problem of inaccurate thermophysical parameters. The training results under different time lengths can predict the changing trends of current and electrode feed rate in advance, enabling early decision-making and prevention. The prediction results under different thermophysical parameter conditions show that the method has good versatility and generalization. This method provides a new technical approach to improve the automation and intelligence level of industrial smelting processes;

[0026] (3) Experimental results show that the Pearson correlation coefficient between the predicted output current and the true value increased from 0.419 to 0.945, and the Pearson correlation coefficient of the output electrode feed rate increased from 0.904 to 0.961. Under the control requirements of the target melting rate and electrode spacing, the fluctuation range of the output current is reduced and the system is more stable.

[0027] (4) The framework of integrating electrode melting dynamics, volumetric Kalman filtering algorithm and recurrent neural network significantly improves the real-time performance, prediction accuracy and robustness of process status. Attached Figure Description

[0028] Figure 1 This is a system structure block diagram of the present invention;

[0029] Figure 2 This is a system workflow diagram of the present invention;

[0030] Figure 3 This is a flowchart of the model training process of the present invention;

[0031] Figure 4 This is a magnified view of the current variation trend in this invention;

[0032] Figure 5 This is a comparison chart of current trends under different time lengths in this invention;

[0033] Figure 6 This is a comparison chart of current trends under different thermophysical parameters in this invention;

[0034] Figure 7 This is a comparison chart of electrode feed rates under different timings in this invention;

[0035] Figure 8 This is a comparison chart of electrode feed rate trends under different thermophysical parameters in this invention. Detailed Implementation

[0036] The present invention will be further described in detail below with reference to embodiments, but the scope of protection of the present invention is not limited thereto.

[0037] This invention relates to a vacuum self-consumption process prediction and control system, configured in conjunction with a vacuum self-consumption furnace, the system comprising:

[0038] A multi-sensor detection unit is used to collect raw process data from the self-vacuum self-consuming furnace;

[0039] A state estimation and feature enhancement module unit is used to receive the original process data and convert it into enhanced feature data;

[0040] A prediction control unit performs online prediction based on the enhanced feature data;

[0041] A control and execution unit is used to output and execute power control and motion control.

[0042] In this invention, a complete industrial-grade closed-loop control system is constructed by integrating a multi-sensor detection module, an embedded physical modeling unit, and an intelligent control actuator. A distributed hardware architecture is adopted, embedding the physical model of electrode melting kinetics into a dedicated processor. Combined with real-time acquisition of process parameters such as current, voltage, and position by the high-precision sensor array of the multi-sensor detection unit, state estimation and feature enhancement are performed using the FPGA's capacitive Kalman filter algorithm. Ultimately, the servo control system of the control actuator precisely adjusts the melting current and electrode feed speed based on the predicted values ​​from the predictive control unit, achieving intelligent and precise control of the vacuum self-consumption process. This effectively solves the problems of response lag, insufficient accuracy, and multi-parameter coupling control inherent in traditional control methods.

[0043] This invention combines a physical model of electrode melting kinetics, a volumetric Kalman filter algorithm, and a recurrent neural network prediction model to form a multi-layered fusion intelligent control framework system, achieving a synergistic effect greater than that of a single method. In the fusion framework of the physical model and the volumetric Kalman filter algorithm, the physical model provides the latter with strong state constraints that conform to metallurgical laws, enabling it to make state estimations in the correct direction even in noisy environments, avoiding the physically unreliable results that may occur with purely data-driven algorithms. In the fusion framework of the volumetric Kalman filter algorithm and the neural network model, the data after filtering and feature enhancement is no longer the original, noisy observation value, but rather high-value information containing the internal state of the process and "purified" by physical laws. This greatly reduces the learning difficulty and data requirements of the neural network model, enabling it to learn the dynamic characteristics of the process more quickly and accurately, and exhibiting excellent generalization ability.

[0044] This hierarchical integration essentially combines human physical cognition of the process (model) with the powerful nonlinear fitting capability of data-driven models, overcoming the limitations of a single method, thereby achieving precise, stable, and adaptive predictive control of the highly nonlinear, time-varying, and large-delay process of vacuum self-consumption.

[0045] The multi-sensor detection unit includes a current sensor for acquiring the actual melting current value of the electrode, a voltage sensor for acquiring the arc voltage, a mass sensor for acquiring the electrode mass, and a position sensor for acquiring the electrode position; the electrode feed speed is obtained based on the change in electrode position within two consecutive sampling periods.

[0046] In this invention, the multi-sensor detection unit is responsible for collecting key process physical parameters from the physical body of the vacuum arc furnace in real time and accurately, providing raw data input for subsequent modules, including but not limited to:

[0047] The current sensor, a non-contact Hall effect current sensor, is installed on the main power supply bus connecting the thyristor rectifier power supply and the electrode clamp, and is used to collect the actual melting current flowing through the consumable electrode in real time. This sensor features a closed-loop measurement structure and low response time, ensuring high-frequency capture of the dynamic current process.

[0048] A mass sensor, a multi-point load cell assembly, which supports the entire electrode lifting mechanism via a load-bearing structure, is used to output an electrical signal in real time that is proportional to the total mass of the consumable electrode, electrode clamp, and lifting mechanism; and a controller whose signal input terminal is connected to the load cell assembly; wherein the controller is programmed to calculate the change in the total mass in real time by receiving the signals from the sensor assembly.

[0049] The voltage sensor is a high-precision isolated voltage sensor, with its measuring terminals connected to the electrode clamp and the conductive plate at the bottom of the crystallizer, respectively, to directly monitor the arc voltage between the electrode tip and the molten pool. Internal voltage reduction and isolation measures are employed to ensure signal accuracy and system safety.

[0050] The position sensor uses an MCS 09F38 servo electrode in vacuum consumable remelting. While driving the electrode to move, the rotary encoder built into the servo motor of the electrode feeding mechanism obtains the angular displacement signal of the motor output shaft. Based on the angular displacement signal and the transmission parameters of the system, the real-time vertical position of the electrode is calculated.

[0051] The controller calculates the real-time electrode feed rate (mm / s) based on the change in electrode position over two consecutive sampling periods.

[0052] The raw process data collected by the vacuum arc remelting furnace has high-frequency noise and includes five variables. In order to track the smelting progress and evaluate the VAR process status during the vacuum arc remelting process, multiple key variables are monitored and recorded as observation vectors, including arc voltage, melting current, electrode position, electrode mass, and electrode feed rate. These process variables will be used as part of the feature quantities to form the input of the recurrent neural network.

[0053] The state estimation and feature enhancement module unit uses the physical model of the electrode melting process as the benchmark model, integrates the volumetric Kalman filter algorithm to filter the raw process data collected by the multi-sensor detection unit, and generates new internal state features.

[0054] The enhanced feature vector is obtained by fusing the output of the state estimation and feature enhancement module units.

[0055] In this invention, the state estimation and feature enhancement module is a dedicated embedded data processing unit integrated in the control cabinet of the VAR furnace. It receives the raw data stream from the sensor module and performs complex calculations in real time to solve the problems of "unmeasurable key states" and "strong noise in sensor signals" during the VAR process. By combining abstract algorithms with physical laws, it derives mathematical models corresponding to these variables based on electrode melting dynamics, accurately captures the dynamic changes in the vacuum arc remelting (VAR) process, and transforms the raw and noisy observation data into enhanced internal state features with richer information and clearer physical meaning, which can provide theoretical support and data foundation for the model.

[0056] The internal state features include electrode thermal boundary layer thickness, efficiency ratio coefficient, current bias, electrode feed rate bias, filling ratio, voltage bias, melting rate, etc. In this embodiment, the number of observation vectors is increased from 5 to 12, so that the recurrent neural network has more learnable features and richer dimensions, thus obtaining more accurate prediction results in applications.

[0057] The electrode mass change rate and electrode gap change rate are obtained based on the electrode melting rate.

[0058] Specifically, the core processor of the state estimation and feature enhancement module unit is an FPGA, such as the National Instruments CompactRIO platform, which has powerful parallel computing capabilities and real-time performance, ensuring the stable operation of the physical model and CKF algorithm; the data acquisition interface is configured with a high-speed, multi-channel analog-to-digital converter, and uses PROFIBUS communication uniformly; the software environment is a real-time operating system (RTOS) that deploys the application program that embeds the physical model and filtering algorithm.

[0059] The baseline model is a one-dimensional heat conduction model with moving boundaries. By combining the basic equations and boundary conditions of the alloy as an electrode, the relationship between the electrode melting rate and the thermal boundary layer change rate is obtained. By embedding a physical model describing the melting process of the 718 alloy electrode, the intrinsic physical laws of the VAR process are established based on the one-dimensional heat conduction equation with moving boundaries. This provides strong prior knowledge and physical constraints for the subsequent filtering algorithm, preventing the model from converging in the wrong direction.

[0060] In model-based control systems, the models used for estimation and control are typically low-order and linear. The essence of controller design lies in inversely solving for the optimal input parameters that meet the target response through the dynamic model. This process can be theoretically expressed as the inverse operation of a system of differential equations. In VAR control problems, the melting rate of the electrodes and the electrode spacing are controlled by adjusting the melting current and electrode feed rate, respectively. This process depends on the electrode temperature distribution and changes in important parameters related to the melting process. To describe this complex process, a simplified one-dimensional heat conduction model with moving boundaries is used. Assuming uniform heating at the lower end of the electrodes, the system modeling equations are as follows:

[0061]

[0062] Where T is the temperature field function with respect to time t and displacement x, and K is the thermal conductivity. C is the electrode density, and C is the specific heat capacity.

[0063] Based on the aforementioned fundamental equations, the relationship between the electrode melting rate and the thermal boundary layer change rate is derived. Because these equations are highly nonlinear, there is no universally applicable solution method; specific numerical methods such as the finite difference method, finite element method, or approximate integral method are typically required. While discrete methods can theoretically achieve arbitrary accuracy, they are complex in practical control applications due to the large number of interrelated equations involved. In contrast, although the accuracy of the approximate integral method may be slightly lower, it is not limited by the aforementioned complexity and is relatively simple to operate.

[0064] Solving this equation yields analytical expressions for the rate of change of the thermal boundary layer and the electrode melting rate:

[0065]

[0066] Among them, C ΔΔ C Δp C sΔ C sp For thermophysical parameters; α r Here, h is the temperature correlation coefficient, Δ is the thermal boundary layer thickness, and h is the temperature correlation coefficient. m It is the enthalpy of melt. Let G be the melt power flux. Based on the analytical expressions of the thermal boundary layer change rate and the electrode melting rate, features that can be used as recurrent neural networks can be derived for subsequent training and prediction. One of these features is the electrode mass change rate, i.e., the melting rate, which can be calculated based on the electrode melting rate. Another important VAR process variable is the electrode gap G, which is also calculated using the electrode melting rate.

[0067] Subsequently, the capacitive Kalman filter algorithm is embedded in the model. Because it approximates the posterior probability density of the Gaussian distribution based on the third-order spherical radial criterion, it does not require linearization, thus having higher estimation accuracy and better numerical stability when dealing with strongly nonlinear problems such as VAR.

[0068] Set the algorithm initialization state vector x 0,0 Error covariance matrix P 0,0 The process noise matrix Q and the observation noise matrix R are used to characterize the process parameters of electrode melting kinetics using a volumetric Kalman filter (VKF) algorithm and make optimal estimates. Historical data is processed to obtain a dataset. The neural network model is based on the dataset processed by the VKF algorithm to obtain the predicted output vector and the predicted state vector. Given the highly nonlinear characteristics of the VAR process, the VKF algorithm is used to more accurately approximate the Gaussian density function, improving prediction accuracy and convergence. Electrode melting kinetics is used as the state model of the VKF algorithm, where the variables measured by sensors are used as the observation vector, and the variables involved in electrode melting kinetics, such as electrode thermal boundary layer thickness, efficiency ratio coefficient, current bias, electrode feed rate bias, filling ratio, voltage bias, melting rate, etc., are used as the state vector. The variables in the state vector and the observation vector are all used as feature quantities, thus successfully expanding the initial 5 feature quantities (electrode gap, electrode mass, current, electrode position, and voltage) to 12 feature quantities.

[0069] In this invention, after processing by the volumetric Kalman filter algorithm based on electrode melting dynamics, not only is the noise of the initial 5 features significantly reduced, but the key thermodynamic and kinetic features of the vacuum self-consuming remelting process are also captured. Finally, an enhanced feature vector of 12 features is output, providing high-quality, high-information-density input for the subsequent predictive control module.

[0070] The predictive control unit predicts the optimal melting current value and electrode feed rate value for the next control cycle.

[0071] The prediction and control unit is a recurrent neural network.

[0072] In this invention, enhanced feature vectors are used to predict the optimal control parameters required to maintain stable melting in the future, namely melting current and electrode feed rate, through a deep learning model. In this embodiment, a large amount of historical melting data is used to train the GRU network offline. The input of the model is the enhanced feature vector, and the output (label) is the corresponding actual control value (melting current and electrode feed rate).

[0073] In actual control, the trained GRU model is deployed on the control system, receives the enhanced feature vector in real time, and predicts the optimal current and electrode feed rate setpoints for the next control cycle.

[0074] In this invention, considering that the datasets calculated by the volumetric Kalman filter algorithm based on electrode melting kinetics have differences in data size and dimensions, they need to be standardized to ensure the comparability between different features, eliminate the influence of dimensions and unfair weight allocation, and accelerate the convergence speed of the learning algorithm. Generally, the Z-Score standardization method is used to process the dataset. The mean of the Z-Score standardized dataset is close to 0, and the standard deviation is close to 1, making it easier to compare and combine different features.

[0075] The interpreter used in this experiment is Python 3.8, PyCharm 2024.2.3, and PyTorch version 2.0.0. The original experimental data comes from the steady-state period of five different furnace runs. For each furnace run, corresponding feature data with a steady-state time step of 29867 seconds is taken. Four furnace runs are used as the training set, and the data from the remaining furnace run is used as the input for the prediction stage to predict the result: electrode feed rate U. ram The correlation between the predicted value and the actual value is evaluated using the Pearson correlation coefficient, with the calculated value ranging from r∈(-1,+1). A larger value indicates a stronger correlation, and it is used to measure the degree of correlation between the predicted value and the actual value.

[0076] The main parameter settings of the recurrent neural network model used are shown in Table 1.

[0077] Table 1. Main parameters of the neural network model

[0078]

[0079] In this invention, K-fold cross-validation is used to perform model evaluation.

[0080] The control execution unit includes a power controller and a motion controller. The power controller controls the optimal melting current value to act between the electrode and the molten pool, and the motion controller controls the motor to act on the electrode so that its movement speed meets the electrode feed speed value.

[0081] In this invention, the control execution unit converts the digital instructions generated by the predictive control module into precise operations on the physical equipment of the VAR furnace, thereby completing the entire control closed loop. Specifically:

[0082] Melting current control: After receiving the current setpoint from the predictive control unit, the module controls a high-power DC power supply through a programmable logic controller. This power supply is based on a silicon controlled rectifier power system and can precisely adjust the output current applied between the electrode and the molten pool.

[0083] Electrode feed speed control: After receiving the electrode feed speed setpoint from the predictive control unit, the module controls the servo motor that drives the electrode lifting arm. Through feedback from the high-precision encoder, it adjusts the output value of the frequency converter to achieve precise control of the electrode descent speed.

[0084] The experimental results of this invention are given below.

[0085] After using the feature enhancement method proposed in this invention (Ad-feature), the GRU model's predicted curves for current and electrode feed rate closely match the measured values. In contrast, the prediction results using only the original sensor data (Feature) show a larger deviation and inaccurate trend capture; as shown in Tables 2-5 below, the significant improvement in the Pearson correlation coefficient quantifies this improvement effect.

[0086] Table 2. Pearson correlation coefficients of current before and after adding features under different models.

[0087]

[0088] Table 3. Pearson correlation coefficients of electrode feed rates before and after adding features under different models.

[0089]

[0090] Robustness verification: such as Figure 4 As shown, in the actual smelting process, when the current changes abruptly (which may be caused by abnormal phenomena such as lumps falling off), the model prediction value proposed in this invention can quickly follow and remain stable, showing good robustness.

[0091] Model generality and scalability verification: such as Figure 6 and Figure 8 As shown, even when the thermophysical parameters in the physical model are artificially amplified by 1.25 times and 1.5 times to simulate inaccurate model parameters, the system can still output prediction results that are highly correlated with the true values, proving the strong anti-interference ability and generalization ability of this method against model uncertainty.

[0092] Table 4. Pearson correlation coefficients for different thermophysical parameters

[0093]

[0094] Forward-looking decision-making ability: Figure 5 and Figure 7The training results of the model on current and electrode feed speed under different timing lengths after adding features; under different timing lengths, the Pearson correlation coefficient still remains at a high level, which indicates that there is a strong linear correlation between the model prediction and the actual measurement value, that is, the model's prediction of the state of the self-consuming furnace a few seconds later is still accurate, which gives the control system more reaction time.

[0095] Table 5. Pearson correlation coefficients under different time lengths

[0096]

[0097] The above embodiments fully demonstrate that the present invention, by constructing a multi-module integrated closed-loop control system, integrates physical models, filtering algorithms and deep learning, effectively solves many problems in the prior art, and provides a complete technical solution for achieving precise, stable and intelligent control of complex industrial processes.

[0098] This invention explicitly proposes a hardware architecture with FPGA (Field Programmable Gate Array) as the core processor. The hardware programmability of FPGA enables it to achieve true parallel computing, and processes multiple calculation steps of physical model and capacitive Kalman filter algorithm in parallel at the hardware level, which greatly shortens the calculation cycle. At the same time, combined with a real-time operating system, it ensures the deterministic execution time of the entire state estimation and prediction task, which meets the stringent requirements of industrial control for high real-time performance.

[0099] In the specific implementation process, the FPGA-based control system and the PLC control system in the industrial field are implemented in parallel with synchronized work cycles and the control subject can be switched to improve the control mode and safety factor. The FPGA control system and the PLC control system use the same furnace body. On the basis of the same furnace body, dedicated sensor detection equipment and control unit processing modules are added to achieve data homogeneity and parallel processing. Based on the output of the predictive control unit, the output is sent to the multi-input / output frequency converter and the thyristor rectifier power supply. Among them, the output of the FPGA-based control system is a digital quantity, which has strong anti-interference ability. Finally, the control execution unit can switch between digital and analog quantities to achieve parallel control.

[0100] The combination of the parallel architecture control system of this invention with traditional control not only improves the smelting stability of vacuum consumable remelting, but also greatly reduces the application risks and costs of new technologies. It also provides unprecedented "flexibility" and redundant backup for the production process, improving the overall safety factor. This is an important creative step for this invention from technical theory to industrial application, and it has significant demonstrative significance for the optimization and improvement of control systems.

[0101] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0102] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0103] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0104] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0105] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention.

[0106] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A vacuum consumable process predictive control system in conjunction with a vacuum consumable furnace, characterized by: The system comprises: a multi-sensor detection unit for collecting raw process data from a vacuum consumable furnace; a state estimation and feature enhancement module unit for receiving the raw process data and converting it into enhanced feature data; a predictive control unit for performing online prediction based on the enhanced feature data; a control execution unit for outputting and executing power control and motion control.

2. A vacuum consumable process predictive control system according to claim 1, wherein: The multi-sensor detection unit comprises a current sensor for collecting actual melting current values of the electrode, a voltage sensor for collecting arc voltage, a mass sensor for collecting electrode mass, and a position sensor for collecting electrode position; The electrode feed speed is obtained based on the change in electrode position in two consecutive sampling periods.

3. A vacuum consumable process predictive control system according to claim 2, wherein: The state estimation and feature enhancement module unit takes the physical model of the electrode melting process as a reference model, fuses the volume Kalman filter algorithm to filter the raw process data collected by the multi-sensor detection unit, and generates new internal state features; The state estimation and feature enhancement module unit outputs the fused enhanced feature vector.

4. A vacuum consumable process predictive control system according to claim 3, wherein: The reference model is a one-dimensional heat conduction model with a moving boundary, which, together with the basic equation and boundary conditions of the heated alloy electrode, obtains the relationship between the electrode melting rate and the heat boundary layer change rate.

5. A vacuum consumable process predictive control system according to claim 4, wherein: Based on the electrode melting rate, the electrode mass change rate and the electrode gap change rate are obtained.

6. A vacuum consumable process predictive control system according to claim 3, wherein: The internal state features include electrode heat boundary layer thickness, efficiency ratio coefficient, current bias, electrode feed speed bias, filling ratio, voltage bias, and melting speed.

7. A vacuum consumable process predictive control system as claimed in claim 1, wherein: The predictive control unit predicts the optimal melting current value and electrode feed speed value in the next control period.

8. A vacuum consumable process predictive control system according to claim 7, wherein: The predictive control unit is a recurrent neural network.

9. A vacuum consumable process predictive control system according to claim 7, wherein: The control execution unit comprises a power controller and a motion controller, the power controller controls the optimal melting current value acting on the electrode and the molten pool, and the motion controller controls the motor acting on the electrode to make its motion speed meet the electrode feed speed value.

Images