Method and system for predicting, controlling and adjusting water flow of self-consuming vacuum furnace for high-temperature alloy

Abstract

The invention belongs to the technical field of high-temperature alloy smelting equipment control, and discloses a method and system for predicting, controlling and adjusting water flow of a consumable vacuum furnace for high-temperature alloy. Aiming at the problems of water flow regulation lag, low prediction precision, large quality fluctuation and high energy consumption in the smelting process of an existing vacuum consumable electrode furnace, the method predicts the water flow in the next five seconds by adopting an LSTM-XGBoost fusion model through multi-source data acquisition and wavelet transform preprocessing, realizes self-adaptive adjustment within one second by combining a fuzzy PID algorithm, and improves the prediction accuracy of the water flow in the next five seconds. The model is reversely optimized based on the cast ingot flaw detection qualification rate and the composition segregation degree; the matching system comprises a data acquisition module, a data processing module, a prediction control module, an execution module, a feedback monitoring module and a software system, and supports off-line operation and multi-mark alloy adaptation. And the flaw detection qualification rate of the high-temperature alloy cast ingot is increased to 97.8%, the cooling water consumption of a single furnace is reduced by 11.8%, the energy consumption of a water pump is reduced by 20.8%, the quality stability of the cast ingot can be effectively guaranteed, the production cost is reduced, and the method is suitable for vacuum consumable remelting production of high-temperature alloys such as GH4169 and the like.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for high-temperature alloy smelting equipment, specifically to a method and system for predicting and controlling the water flow of a vacuum consumable furnace for high-temperature alloys. Background Technology

[0002] Vacuum arc remelting furnaces for high-temperature alloys achieve melting and remelting of high-temperature alloys through arc discharge between electrodes and ingots. The flow stability of its cooling water system directly determines the ingot cooling rate, solidification uniformity, and surface quality. Currently, the industry suffers from the following core technical deficiencies in water flow control:

[0003] 1. Significant lag in regulation: Existing technologies mostly adopt a passive regulation mode of "real-time monitoring - deviation correction," relying on PID algorithms to correct for existing flow deviations. When parameters such as furnace temperature (e.g., sudden changes in crystallizer wall temperature ±20℃) and electrode melting rate (e.g., electrode pull rod speed increasing from 3mm / min to 5mm / min) fluctuate drastically, the regulation lag time is generally ≥3 seconds, which can easily lead to local overheating of the crystallizer, causing surface cracks or internal porosity in the ingot, resulting in a yield rate of only 80%-85%.

[0004] 2. Poor adaptability of prediction models: Traditional control relies on a single machine learning model (such as LSTM) or empirical formula, which cannot simultaneously handle "continuous fluctuations in time series data" (such as gradual changes in current over time) and "nonlinear sudden disturbances" (such as temperature spikes caused by arc instability). The prediction error often exceeds 12%, leading to over- or under-regulation.

[0005] 3. Lack of quality closed loop: The control logic is only related to real-time process parameters and is not included in the final quality feedback of the ingot (such as flaw detection results and composition segregation). It is impossible to dynamically optimize the flow parameters for different alloy grades or ingot specifications, resulting in a quality fluctuation of ±15% for the same batch of products.

[0006] 4. Inefficient energy consumption control: To avoid insufficient cooling, an "over-cooling" strategy is often adopted, with the cooling water time for a single furnace exceeding the actual demand by 15%-20%, resulting in wasted electricity. Summary of the Invention

[0007] This invention aims to solve the problems of lagging water flow regulation, low prediction accuracy, weak quality correlation and high energy consumption in vacuum arc furnaces. By constructing a technical system of "multi-source data fusion prediction model + fuzzy PID adaptive control + quality closed-loop optimization", it can achieve 5-second advance prediction and dynamic correction of water flow, improve the yield of high-temperature alloy ingots to over 95%, and reduce energy consumption per furnace by 10%-15%.

[0008] To achieve the above objectives, the present invention provides the following technical solution: The technical solution of the present invention consists of two parts: a predictive control method and a supporting system, which work together to achieve full-link regulation.

[0009] (I) Predictive Control and Regulation Method for Water Flow Rate in Vacuum Arsenic Furnaces for High-Temperature Alloys

[0010] This method comprises the following five core steps, forming a closed loop of "data input - predictive output - control execution - quality feedback":

[0011] Real-time acquisition of multi-source parameters

[0012] A distributed sensor network was used to collect six key parameters of the vacuum arc furnace, with a sampling frequency of 1Hz to ensure data timeliness.

[0013] Furnace internal parameters: crystallizer wall temperature (6 PT100 sensors evenly arranged, measurement range 0-500℃, accuracy ±0.5℃), electrode area arc temperature (OS137 infrared thermometer, measurement range 800-1800℃, accuracy ±2%), furnace vacuum degree (GP-100 vacuum gauge, measurement range 10-3-10Pa).

[0014] Electrical parameters: Electrode current (Hall sensor CSM025, 0-27kA, accuracy ±0.2%), Electrode voltage (voltage divider sensor LV28-P, 0-100V, accuracy ±0.5%);

[0015] Cooling water system parameters: Inlet temperature (platinum resistance sensor, 0-100℃), outlet temperature (same as above), inlet pressure (diffused silicon sensor MPX5700, 0-2MPa), outlet pressure (same as above), current water flow rate (electromagnetic flowmeter LDG-MIK, 0-50m³ / h). 3 / h, accuracy ±0.5%;

[0016] Process parameters: electrode rod pulling speed (encoder E6B2-CWZ6C, 0-5mm / min), alloy grade (preset parameters, such as GH4169), ingot specifications (diameter × length, preset parameters).

[0017] Data preprocessing

[0018] The collected data undergoes a three-stage purification process to eliminate interference and standardize it:

[0019] Noise filtering: The arc spike noise (amplitude > 3 times standard deviation) in temperature and current signals is hard-thresholded using the db4 wavelet basis (3-level decomposition), improving the signal-to-noise ratio to ≥30dB;

[0020] Missing value completion: For missing data (≤0.5%) caused by sensor interruption, linear interpolation is used to complete the missing data to ensure temporal continuity;

[0021] Normalization: Map all parameters to the [0,1] interval through standardization (formula: This eliminates the impact of magnitude differences on the model.

[0022] Fusion prediction model construction and computation

[0023] A "LSTM-XGBoost two-layer fusion model" is constructed to predict water flow demand in the next 5 seconds, overcoming the limitations of a single model:

[0024] The underlying LSTM module processes temporal features (such as the continuous changing trends of temperature and current), sets up 3 hidden layers (64, 32, and 16 neurons), and uses the ReLU activation function to capture the dynamic correlation of parameters over time.

[0025] Top-level XGBoost module: Integrates nonlinear features (such as the mapping relationship between alloy grade and flow rate, and the impact of sudden changes in vacuum degree), sets the number of trees to 100, the learning rate to 0.1, and the maximum depth to 6, and handles discrete interference factors;

[0026] Model training: The model is trained using historical production data (complete parameters for ≥100 batches). The root mean square error (RMSE) is minimized based on the Adam optimizer (learning rate 0.001, β1 = 0.9, β2 = 0.999). Hyperparameters are optimized through 5-fold cross-validation to bring the prediction error to ≤5%.

[0027] Fuzzy PID adaptive closed-loop control regulation

[0028] Based on the deviation between the predicted traffic demand and the real-time traffic, dynamic adjustment commands are output:

[0029] Deviation calculation: Define deviation e = actual flow - predicted demand, deviation rate of change.

[0030] Fuzzy reasoning: Divide e and Δe into 7 fuzzy subsets (negative large, negative medium, negative small, zero, positive small, positive medium, positive large), establish 49 fuzzy rules (such as "e = positive large and Δe = negative small → output = positive large"), and obtain the PID parameter correction amount by defuzzifying through the centroid method;

[0031] Dynamic parameter correction: Real-time adjustment of the PID proportional coefficient (K) P ), integral coefficient (K) i ), differential coefficients (K) d ), where K P Adjustment range 0.5-2.0, K iAdjustment range 0.01-0.1, K d Adjustment range: 0.1-1.0; Execution output: Convert the corrected control quantity into a 4-20mA analog signal to adjust the opening of the electric regulating valve (accuracy ±0.5%) or the speed of the variable frequency water pump (0-50Hz), so that the actual flow rate approaches the predicted value within 1 second, with steady-state fluctuation ≤±3%.

[0032] Quality feedback model optimization

[0033] After each heat cycle, the ingot quality data is fed back into the model for continuous optimization.

[0034] Quality data acquisition: The flaw detection pass rate is obtained through ultrasonic flaw detector, the compositional segregation degree (such as Nb element segregation) is obtained through spectrometer, and the grain size grade is obtained through metallographic microscope.

[0035] Model correction trigger condition: If the flaw detection pass rate is <95% or the segregation is >0.8%, the model weight correction will be initiated;

[0036] The algorithm was modified by using gradient descent to fine-tune the connection weights of the LSTM-XGBoost model (learning rate 0.0001) so that the next batch of traffic predictions better matches the quality requirements, forming a closed loop of "process-quality-control".

[0037] (II) Predictive Control System for Water Flow in Vacuum Arsenic Furnace for High-Temperature Alloys

[0038] This system serves as the hardware platform for the aforementioned method, comprising five major functional modules and supporting software, with the following structure: Figure 1 As shown:

[0039] Data acquisition module

[0040] Includes 12 sensors and signal conditioning unit:

[0041] Temperature sensors: PT100 type (crystallizer wall temperature), infrared thermometer (model OS137);

[0042] Electrical parameter sensors: 1 Hall current sensor (model CSM025), 1 voltage divider sensor (model LV28-P);

[0043] Water system sensors: 1 electromagnetic flow meter (model LDG-MIK), 2 diffused silicon pressure sensors (model MPX5700), and 2 platinum resistance temperature sensors (inlet and outlet);

[0044] Signal conditioning: An isolation amplifier (model ISO124) is used to eliminate electromagnetic interference and convert the sensor signal into a 0-5V standard signal with a sampling accuracy of 16 bits to ensure data accuracy.

[0045] Data processing module

[0046] Built on an edge computing gateway (model ECG-100), it achieves the following:

[0047] Real-time data reception: Sensor data is received in real time via Modbus-RTU protocol (baud rate 9600bps);

[0048] Preprocessing: Built-in wavelet transform, interpolation and normalization algorithms, processing delay ≤50ms;

[0049] Data forwarding: Preprocessed data is synchronized to the predictive control module (real-time) and the database (storage) via Ethernet.

[0050] Predictive control module

[0051] The core is an embedded industrial computer (model IPC-610L, equipped with an NVIDIA Jetson TX2 GPU):

[0052] Model execution: Load the LSTM-XGBoost fusion model, single inference time ≤200ms, supports offline execution;

[0053] Control logic: Execute fuzzy PID algorithm, output 4-20mA analog control signal, control cycle 100ms;

[0054] Fault diagnosis: Real-time monitoring of sensor disconnection (signal over range) and actuator jamming (command and feedback deviation > 10%) triggers audible and visual alarms (alarm sound pressure ≥ 85dB, red light flashing).

[0055] Execution module

[0056] It includes two types of actuators to regulate water flow:

[0057] Electric regulating valve: Model ZDLP-16C, nominal diameter DN50, regulating range 0-100%, response time ≤0.8 seconds;

[0058] Variable frequency water pump: Model ISG50-160, power 7.5kW, speed adjustment range 0-2900r / min, flow rate adjustment accuracy ±1%.

[0059] Feedback monitoring module

[0060] Connect to quality inspection equipment and transmit quality data:

[0061] Interface: Communicates with ultrasonic flaw detectors and spectrometers via OPC UA protocol (version 1.04), with data transmission rate ≥1Mbps and latency ≤50ms;

[0062] Data analysis: Extract key indicators such as flaw detection pass rate and segregation, format them, and upload them to the predictive control module.

[0063] Supporting software system

[0064] Developed based on Windows Server 2024, it includes 4 subsystems:

[0065] Data management subsystem: It adopts MySQL 8.0 database, stores ≥10 years of historical data (based on 200 furnaces / year), supports multi-dimensional queries by furnace ID and alloy grade, and the storage size of a single record is ≤128 bytes;

[0066] Model computation subsystem: provides model training (supports incremental learning), parameter tuning, and version switching functions, and has built-in model performance evaluation metrics (RMSE, MAE);

[0067] Control and regulation subsystem: Visual configuration of prediction time window (default 5 seconds), fuzzy rule table, PID parameter range, and supports manual / automatic control switching;

[0068] Monitoring subsystem: Real-time display of parameter curves (temperature, flow rate, current), comparison of predicted and actual values, alarm information, with an interface refresh rate of 1Hz.

[0069] Compared with the prior art, the beneficial effects of the present invention are:

[0070] 1. High prediction accuracy: The water flow prediction error of the LSTM-XGBoost fusion model is ≤5%, which is more than 58% lower than that of the single LSTM model (error 12%), avoiding adjustment deviations caused by inaccurate prediction;

[0071] 2. Fast response speed: The end-to-end latency from data acquisition to actuator action is ≤1 second, which is 200% faster than traditional PID regulation (3-second delay), and can quickly respond to sudden changes in furnace parameters;

[0072] 3. Improved quality stability: Through closed-loop optimization of quality feedback, the pass rate of flaw detection of high-temperature alloy ingots increased from 82.3% to 97.8% (+15.5 percentage points), the compositional segregation decreased from 0.92% to 0.51% (-44.6%), and the grain size was refined to grade 1.6 (ASTM standard).

[0073] 4. Significantly reduced energy consumption: Precise control avoids excessive cooling, reducing cooling water consumption per furnace cycle from 85m³ / h. 3 Reduced to 75m 3 (-11.8%), water pump energy consumption decreased from 120kWh to 95kWh (-20.8%), resulting in cost savings;

[0074] 5. High adaptability: Supports multiple alloy grades such as GH4169, GH718, and GH4151. It can adaptively adjust the model parameters to suit different ingot specifications (Φ300mm-Φ800mm) without reprogramming. Attached Figure Description

[0075] Figure 1 A schematic diagram of the overall framework of a water flow prediction and control system for a vacuum arc furnace used in high-temperature alloys;

[0076] Figure 2 This is a flowchart of a fusion prediction model (LSTM-XGBoost) for a water flow prediction and control system of a vacuum arc furnace for high-temperature alloys.

[0077] Figure 3 A fuzzy PID control logic diagram for a water flow prediction and control system of a vacuum arc furnace for high-temperature alloys;

[0078] Figure 4 This is a comparison chart showing the water flow control effect of GH4169 alloy in the examples. Detailed Implementation

[0079] 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.

[0080] Please see Figures 1-4 This invention provides a technical solution: taking GH4169 high-temperature alloy vacuum arc remelting (ingot specifications Φ508mm×2000mm) as an example, the implementation process of this invention is described in detail:

[0081] 1. System Setup

[0082] Data acquisition module: Six PT100 sensors (measurement range 0-500℃) are evenly arranged around the circumference of the crystallizer; a CSM025 current sensor (0-27kA) is installed at the water-cooled cable; and an LDG-MIK electromagnetic flow meter (0-50mA) is installed on the cooling water pipe. 3 / h);

[0083] Predictive control module: It adopts an IPC-610L industrial computer, equipped with a Jetson TX2 GPU, and a pre-trained LSTM-XGBoost model (3 LSTM hidden layers and 100 XGBoost trees);

[0084] Execution module: The main cooling water pipeline is equipped with a ZDLP-16C electric regulating valve and an ISG50-160 variable frequency water pump;

[0085] Feedback monitoring module: Connects to ultrasonic flaw detector and spectrometer.

[0086] 2. Operating Procedures

[0087] Startup and parameter settings: Open the system software, select the "GH4169" alloy grade, set the prediction time window to 5 seconds and the flow fluctuation threshold to ±3%;

[0088] Data acquisition and preprocessing: The system acquires parameters such as temperature, current, and flow rate at a frequency of 1Hz, filters out arc noise through wavelet transform, and completes missing data through linear interpolation;

[0089] Flow forecast: The fusion model performs calculations once per second, outputting the water flow demand value for the next 5 seconds (e.g., when the electrode current increases from 9kA to 10kA, the predicted flow rate increases from 25m³ / h). 3 / h increased to 28m 3 / h);

[0090] Closed-loop control: Fuzzy PID algorithm based on "actual flow rate (25m³)" 3 / h) - Forecasted demand (28m) 3 The deviation of " / h)" causes the output control signal to increase the opening of the electric regulating valve from 50% to 58%, and the flow rate reaches 28m³ / h within 300ms. 3 / h;

[0091] Quality feedback: After the furnace cycle, ultrasonic testing showed that the ingot pass rate was 98% (82% with traditional PID control), and spectral analysis showed a segregation rate of 0.5% (0.9% with traditional control). The system automatically fine-tuned the model weights and optimized the control parameters for the next batch.

[0092] 3. Implementation Results

[0093] Water flow fluctuation range: reduced from ±15% of traditional PID to ±3%;

[0094] Ingot crack rate: decreased from 8% to 1.2%;

[0095] Production time per furnace: reduced from 48 hours to 42 hours (due to stable cooling, the electrode rod pulling speed can be increased by 10%).

[0096] The advantages of this invention are mainly reflected in the following aspects:

[0097] Compared with the prior art, this invention has the following four core technical advantages:

[0098] 1. High prediction accuracy: The water flow prediction error of the LSTM-XGBoost fusion model is ≤5%, which is more than 58% lower than that of the single LSTM model (error 12%), avoiding adjustment deviations caused by inaccurate prediction;

[0099] 2. Fast response speed: The end-to-end latency from data acquisition to actuator action is ≤1 second, which is 200% faster than traditional PID regulation (3-second delay), and can quickly respond to sudden changes in furnace parameters;

[0100] 3. Improved quality stability: Through closed-loop optimization of quality feedback, the pass rate of flaw detection of high-temperature alloy ingots increased from 82.3% to 97.8% (+15.5 percentage points), the compositional segregation decreased from 0.92% to 0.51% (-44.6%), and the grain size was refined to grade 1.6 (ASTM standard).

[0101] 4. Significantly reduced energy consumption: Precise control avoids excessive cooling, reducing cooling water consumption per furnace cycle from 85m³ / h. 3 Reduced to 75m 3 (-11.8%), water pump energy consumption decreased from 120kWh to 95kWh (-20.8%), saving 2,850 yuan in costs for 10 furnace cycles.

[0102] 5. High adaptability: Supports multiple alloy grades such as GH4169, GH718, and GH4151. It can adaptively adjust the model parameters to suit different ingot specifications (Φ300mm-Φ800mm) without reprogramming.

[0103] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predicting and controlling the water flow rate of a vacuum arc furnace for high-temperature alloys, characterized in that, Includes the following steps: Step 1) Multi-source data acquisition: Real-time acquisition of furnace temperature, electrode current and voltage, cooling water parameters, ingot casting speed and alloy grade of vacuum arc furnace, with a sampling frequency of 1Hz; Step 2) Data preprocessing: Wavelet transform is used for noise reduction, linear interpolation for completion, and normalization of the collected data. Step 3) Fusion prediction model calculation: Predict the water flow demand value for the next 5 seconds based on the LSTM-XGBoost fusion model, with a prediction error ≤5%; Step 4) Adaptive closed-loop control: The fuzzy PID algorithm is used to dynamically adjust the opening of the electric regulating valve or the speed of the variable frequency water pump according to the flow deviation, with an adjustment response delay of ≤1 second; Step 5: Quality feedback optimization, based on the ingot flaw detection pass rate and composition segregation degree, correct the fusion model parameters.

2. The method according to claim 1, characterized in that, The cooling water parameters in step 1 include inlet temperature, outlet temperature, inlet pressure, outlet pressure, and current water flow rate, which are collected using a platinum resistance sensor, a diffused silicon pressure sensor, and an electromagnetic flow meter, respectively.

3. The method according to claim 1, characterized in that, In step 3, the LSTM module has 3 hidden layers with 64, 32 and 16 neurons respectively; the XGBoost module has 100 trees and a learning rate of 0.1, and the hyperparameters are optimized through 5-fold cross-validation.

4. The method according to claim 3, characterized in that, In step 3, the fusion model is trained using the Adam optimizer with a learning rate of 0.001, β1 = 0.9, β2 = 0.999, and the loss function is the root mean square error. The validation set error converges to ≤4.5%.

5. The method according to claim 1, characterized in that, The wavelet transform in step 2 uses the db4 wavelet basis and has a decomposition layer of 3. Hard thresholding is applied to spike noise with amplitude > 3 times the standard deviation in temperature and current signals.

6. The method according to claim 1, characterized in that, The fuzzy PID algorithm in step 4 contains 49 fuzzy rules, which divide the deviation and the rate of change of deviation into 7 fuzzy subsets. The output is defuzzified by the centroid method and converted into a 4-20mA control signal.

7. A water flow prediction and control system for a vacuum arc furnace for high-temperature alloys, implementing the method of any one of claims 1-6, characterized in that, include: Data acquisition module: includes temperature, current, and flow sensors and a signal conditioning unit; Data processing module: Implements data cleaning and forwarding based on edge computing gateway; Predictive control module: an industrial control computer equipped with a GPU, running an LSTM-XGBoost model and a fuzzy PID algorithm; Execution modules: electric regulating valve and variable frequency water pump; Feedback monitoring module: Connects to ultrasonic flaw detector and spectrometer; Software system: includes data management, model calculation, control and regulation and monitoring subsystems.

8. The system according to claim 7, characterized in that, The temperature sensors in the data acquisition module include a PT100 type and an infrared thermometer, the current sensor is a Hall effect type, and the flow sensor is an electromagnetic flow meter.

9. The system according to claim 7, characterized in that, The data management subsystem of the software system uses a MySQL 8.0 database, which supports multi-dimensional indexing by process card number and alloy grade, and the space occupied by a single data storage is ≤128 bytes.

10. The system according to claim 7, characterized in that, The predictive control module uses an NVIDIA Jetson TX2 GPU, with a model inference time of ≤200ms / inference and supports offline deployment; the execution module has an electric regulating valve response time of ≤0.8 seconds and a variable frequency water pump power of 7.5kW. The feedback monitoring module communicates with the quality inspection equipment via OPC UA protocol version 1.04, with a data transmission rate ≥1Mbps and a communication delay ≤50ms.

Images