Alloy blank continuous filling control method based on multi-source data fusion

Abstract

The invention discloses an alloy blank continuous filling control method based on multi-source data fusion, and the method comprises the steps: collecting multi-source heterogeneous data in real time, and driving a dynamic digital twin model to carry out online updating and calibration, so as to map the process state of a filling process in real time; a hybrid intelligent driving algorithm is adopted, a prediction result of the dynamic digital twinborn model is combined with error compensation of a space-time attention long-term and short-term memory network, multi-target rolling optimization is conducted through a model prediction control algorithm, and a future optimal alloy filling rate curve is generated; converting the optimal rate curve into a synchronous control instruction for a plurality of execution mechanisms based on a cooperative control strategy mapped by a virtual main shaft and an electronic cam; the method further comprises adaptive replay learning and real-time filling quality evaluation functions. According to the method, through deep fusion of data driving and model driving, the alloy filling precision, efficiency and process stability are improved, and the method has wide application prospects.

Description

Technical Field

[0001] This invention belongs to the field of industrial process control and intelligent manufacturing technology, and specifically discloses a method for continuous loading control of alloy billets based on multi-source data fusion. Background Technology

[0002] In metallurgy, casting, and other fields, continuous loading of alloy billets is a crucial process, and its control precision directly determines the compositional uniformity, microstructure, and mechanical properties of the final product. With the development of Industry 4.0 and intelligent manufacturing, unprecedented challenges have been posed to the real-time performance, accuracy, and intelligence level of the alloy loading process.

[0003] Traditional alloy billet loading control methods often rely on operator experience or are based on simple PID (proportional-integral-derivative) control algorithms. These methods typically suffer from the following significant drawbacks:

[0004] Traditional methods often rely on single or a few sensor data points (such as weight signals from weighing sensors) for feedback control, failing to comprehensively and in real-time reflect the complex and ever-changing state of the smelting process. This includes factors such as the temperature distribution within the molten pool, element loss, melt flow, and slag phase state. This lack of information weakens the control system's ability to perceive process changes.

[0005] Static models are rigid and lack adaptability: Most existing control models are static mechanistic models based on offline experiments or theoretical derivations. However, the actual smelting process is affected by a variety of dynamic factors such as batch differences in raw materials, equipment aging, and environmental disturbances. Static models cannot accurately describe these complex and nonlinear dynamic behaviors, resulting in large deviations between model predictions and actual processes, and making it difficult to guarantee control accuracy.

[0006] Lagging control strategies and insufficient optimization capabilities: Traditional feedback control is "retroactive adjustment," meaning it corrects errors only after deviations are detected, resulting in inherent lag. For variables requiring forward-looking planning, such as filling rate, simple feedback control cannot perform rolling optimization based on future process requirements, making it difficult to achieve globally optimal control effects. This may lead to excessive component levels, increased energy consumption, or low production efficiency.

[0007] The difficulty of coordinating multiple actuators: Continuous filling systems typically include multiple actuators, such as feed belts, vibrating feeders, and gate valves. The movements of these mechanisms require a high degree of synchronization and precise coordination to achieve the set filling rate curve. Traditional control methods, when dealing with multi-axis synchronization and complex trajectory planning, often rely on complex hardware cams or cumbersome PLC (Programmable Logic Controller) programs, resulting in poor flexibility, difficult debugging, and difficulty in quickly responding to process changes.

[0008] Therefore, it is necessary to invent a continuous loading control method for alloy billets based on multi-source data fusion to solve the above problems. Summary of the Invention

[0009] To overcome the aforementioned deficiencies in the prior art, this invention provides a continuous alloy billet loading control method based on multi-source data fusion. This method involves real-time acquisition of multi-source heterogeneous data to drive online updates and calibration of a dynamic digital twin model, thereby mapping the process status of the loading process in real time. A hybrid intelligent driving algorithm is employed, combining the prediction results of the dynamic digital twin model with error compensation from a spatiotemporal attention long short-term memory network. Multi-objective rolling optimization is then performed using a model predictive control algorithm to generate the optimal alloy loading rate curve for the future. A collaborative control strategy based on virtual spindle and electronic cam mapping transforms the optimal rate curve into synchronous control commands for multiple actuators. The method also includes adaptive replay learning and real-time loading quality assessment functions, effectively solving the problems mentioned in the background art.

[0010] To achieve the above objectives, the present invention provides the following technical solution: a continuous loading control method for alloy billets based on multi-source data fusion, specifically including the following steps:

[0011] S1. Real-time acquisition of multi-source heterogeneous data during the continuous loading process of alloy billets;

[0012] S2. Based on the multi-source heterogeneous data, drive a dynamic digital twin model to perform online updates and calibrations, so as to map and output the process status of the filling process in real time.

[0013] S3. Based on the real-time process status output by the dynamic digital twin model, a hybrid intelligent drive algorithm is used for rolling optimization to generate the optimal alloy loading rate curve for the future.

[0014] S4. Based on the optimal alloy loading rate curve, a collaborative control strategy based on virtual spindle and electronic cam mapping is used to generate synchronous control commands for at least two actuators to complete precise continuous loading.

[0015] Preferably, in step S1, the multi-source heterogeneous data includes:

[0016] Raw material property data: composition, moisture content, and particle size distribution of the alloy material;

[0017] Process status data: temperature field of melting chamber, melt surface image and intermittent component spectrum signal;

[0018] Equipment operating data: feeding mechanism speed, gate opening and weighing sensor readings.

[0019] Preferably, in step S1, the real-time acquisition of the process status data specifically includes:

[0020] Multi-point temperature field data of the melting chamber were collected using an infrared thermal imager;

[0021] The compositional spectral signals of elements on the surface of the molten pool were intermittently acquired using a laser-induced breakdown spectrometer.

[0022] Visual image data of the melt surface is acquired using a high-speed industrial camera to analyze molten pool fluctuations and slag phase state.

[0023] Preferably, in step S2, the dynamic digital twin model is a lightweight coupling model, and its construction and operation include:

[0024] A mechanistic model framework based on material balance and reaction kinetics was constructed to simulate the physical laws governing the loading process and the molten pool response.

[0025] A real-time calibration module based on a variational autoencoder is introduced, which takes the multi-source heterogeneous data as input, dynamically infers and compensates for the system errors of the mechanistic model skeleton, and realizes online calibration of the model output.

[0026] Preferably, the system error includes dynamic changes in element burn-off rate, fluctuations in heat loss, and real-time deviations in raw material reaction efficiency; the variational autoencoder encodes the latent feature distribution of the multi-source heterogeneous data and fuses it with the mechanistic model prediction in the latent space to output a calibrated state prediction.

[0027] Preferably, in step S3, the execution of the hybrid intelligent driving algorithm includes:

[0028] Using the dynamic digital twin model as the basic predictor, trend predictions that conform to physical laws are provided;

[0029] A spatiotemporal attention long short-term memory network is used as an intelligent compensator to learn the nonlinear relationship between multi-source data sequences in historical data and the prediction error of the basic predictor, and to perform dynamic error compensation for the current prediction result.

[0030] A model predictive control algorithm is adopted, with cost, component hit rate and process stability as multiple objectives. The compensated prediction results are rolled to optimize and solve, and the optimal alloy loading rate curve is generated.

[0031] Preferably, in the optimization objective function of the model predictive control algorithm, the process stability is quantified and constrained by the second derivative of the loading rate curve to minimize the acceleration change of the actuator and ensure smooth loading action.

[0032] Preferably, in step S4, the specific implementation steps of the collaborative control strategy based on virtual spindle and electronic cam mapping include:

[0033] S401: Integrate the optimal alloy loading rate curve into a virtual spindle displacement curve based on time or process progress.

[0034] S402: Predefine the cam curve mapping relationship between each actuator and the virtual spindle displacement curve;

[0035] S403: During the control cycle, each actuator calculates and executes its own motion control command synchronously with the real-time displacement of the virtual spindle according to its own cam curve mapping relationship.

[0036] Preferably, the method further includes an adaptive replay and learning step:

[0037] Record the virtual spindle displacement curve and the actual control parameters of each actuator during the successful production process to form a standard process package;

[0038] When replaying the standard process package, the actual feedback value and command value of each actuator are compared in real time, and instantaneous fine-tuning is performed through a feedforward compensation algorithm. The basis for the fine-tuning includes the real-time deviation of the weighing sensor and the abnormal signal of the vibration sensor.

[0039] Preferably, the method further includes a real-time filling quality assessment step:

[0040] Based on the real-time mapping of the process status in step S2, especially the dynamic deviation between the predicted and target values ​​of key components in the molten pool, the overall quality score of the current filling process is calculated.

[0041] When the overall quality score is lower than a preset threshold, the hybrid intelligent drive algorithm is triggered to start the re-optimization process and correct the subsequent loading rate curve.

[0042] The technical effects and advantages of this invention are as follows:

[0043] 1. It covers three types of multi-source heterogeneous data: raw material properties, process status, and equipment operation. Combined with the precise acquisition by professional equipment such as infrared thermal imagers and laser-induced breakdown spectrometers, it has constructed a panoramic data base for the filling process, avoiding the limitations of a single data dimension.

[0044] The data acquisition takes into account both real-time performance and multidimensionality. It can capture dynamic processes such as temperature field and molten pool fluctuation in the melting chamber, and accurately obtain key parameters such as raw material composition and equipment speed. This provides high-quality data support for subsequent models and algorithms, and solves the problem of data fragmentation in traditional control.

[0045] 2. The lightweight digital twin model, which couples the mechanistic model with the variational autoencoder, not only follows the physical laws of material balance and reaction kinetics, but also dynamically compensates for systematic errors such as element burn-off rate and heat loss fluctuations. Its prediction accuracy is significantly better than that of a single mechanistic model or a data-driven model.

[0046] The model supports online updates and calibration. By fusing latent features from multi-source data with latent space predictions of mechanisms, it achieves real-time and accurate mapping of process states, providing a reliable state basis for subsequent optimization control and adapting to the dynamic characteristics of industrial production.

[0047] 3. The hybrid intelligent driving algorithm integrates physical mechanisms and AI technology, uses a digital twin model as the basic predictor, and uses a spatiotemporal attention LSTM to compensate for nonlinear errors, thus balancing the physical robustness and dynamic adaptability of the prediction.

[0048] The model predictive control algorithm takes cost, component hit rate and process stability as multiple objectives. It ensures smooth operation by constraining the second derivative of the loading rate curve, which improves product quality consistency and reduces actuator losses, thus overcoming the limitations of single-objective optimization in traditional control.

[0049] 4. Based on the collaborative control strategy of virtual spindle and electronic cam mapping, the optimal speed curve is transformed into a displacement curve with a unified reference, realizing precise synchronous control of multiple actuators. The synchronization accuracy far exceeds that of traditional mechanical cam or independent control methods.

[0050] The newly added adaptive replay and learning functions, through the reuse of standard process packages and real-time feedforward compensation, can not only quickly replicate successful production experience, but also cope with sudden situations such as weighing deviation and equipment vibration, thus improving the system's anti-interference capability. Attached Figure Description

[0051] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0052] Figure 1 This is a schematic diagram of the overall structure of the present invention.

[0053] Figure 2 This is a diagram illustrating the collaborative control strategy of virtual spindle and electronic cam mapping in this invention.

[0054] Figure 3 This is a diagram illustrating the adaptive playback and learning steps of the present invention.

[0055] Figure 4 This is a diagram illustrating the real-time filling quality assessment steps of the present invention. Detailed Implementation

[0056] 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.

[0057] This invention provides, for example Figure 1 The continuous loading control method for alloy billets based on multi-source data fusion, as shown, includes the following steps:

[0058] S1. Real-time acquisition of multi-source heterogeneous data during the continuous loading process of alloy billets;

[0059] Furthermore, in the above technical solution, in step S1, the multi-source heterogeneous data includes:

[0060] Raw material property data: composition, moisture content, and particle size distribution of the alloy material;

[0061] Process status data: temperature field of melting chamber, melt surface image and intermittent component spectrum signal;

[0062] Equipment operating data: feeding mechanism speed, gate opening and weighing sensor readings.

[0063] It should be further noted that the alloy composition was measured using an X-ray fluorescence spectrometer (detection accuracy ±0.01%), covering key alloying elements such as C, Si, Mn, and Cr; humidity was collected using an infrared moisture analyzer (measurement range 0-20%, accuracy ±0.1%); and particle size distribution was collected using a laser particle size analyzer (measurement range 0.1-1000μm, repeatability ±2%).

[0064] Batch testing is conducted before raw materials enter the warehouse. Five samples are randomly selected from each batch of raw materials, with a sample weight of ≥50g. The average value of the test results is taken as the attribute data of the batch of raw materials. The data is uploaded to the industrial control system in real time. If the moisture content of a batch of raw materials exceeds 5% or the standard deviation of particle size distribution is >100μm, a pretreatment warning for raw materials (drying or sieving) is triggered.

[0065] Furthermore, in the above technical solution, step S1, specifically including the real-time acquisition of the process status data, includes:

[0066] Multi-point temperature field data of the melting chamber were collected using an infrared thermal imager;

[0067] The compositional spectral signals of elements on the surface of the molten pool were intermittently acquired using a laser-induced breakdown spectrometer.

[0068] Visual image data of the melt surface is acquired using a high-speed industrial camera to analyze molten pool fluctuations and slag phase state.

[0069] It should be further explained that the multi-point temperature field data is obtained by symmetrically arranging three infrared thermal imagers (temperature range 500-1800℃, accuracy ±5℃) on the top of the melting chamber, with the lenses facing the surface of the molten pool, at a frame rate of 10 frames / second, and using a 5×5 grid to divide the measuring points (a total of 25 measuring points), and outputting the temperature value of each measuring point in real time; abnormal measuring points are removed by median filtering (the temperature difference with the surrounding 4 measuring points > 50℃ is judged as abnormal), and temperature field thermogram data is generated.

[0070] The compositional spectral signals were acquired intermittently using a laser-induced breakdown spectrometer. The probe was installed 1.5m above the molten pool at a 45° angle to the liquid surface. The samples were acquired every 30 seconds, with 3 consecutive spectral samples acquired each time (sampling interval of 1 second). The maximum and minimum values ​​were removed, and the average spectrum was taken.

[0071] Visual image data is captured in real time using a high-speed industrial camera (1920×1080 resolution, 200 frames / second) equipped with a high-temperature resistant lens and infrared filter. The molten pool surface image is then captured by Gaussian filtering for noise reduction and adaptive threshold segmentation to extract the molten area. The fluctuation amplitude of the molten pool (the difference between the highest and lowest points of the liquid surface) and the area ratio of the slag phase are then calculated.

[0072] Furthermore, the rotational speed of the feeding mechanism is acquired by an incremental encoder (1024 lines resolution) at a sampling frequency of 50Hz, and the actual rotational speed (unit: r / min) is output after frequency doubling.

[0073] The gate opening is acquired using a pull-wire displacement sensor (measurement range 0-500mm, accuracy ±0.1mm), sampling frequency 20Hz, and the opening percentage is directly output (0-100%).

[0074] The weighing sensor uses a tensile sensor (range 0-500kg, accuracy ±0.2%FS) with a sampling frequency of 100Hz. Vibration noise is eliminated through Kalman filtering, and the real-time material weight is output.

[0075] The vibration sensor uses a piezoelectric accelerometer (measurement range 0-5g, accuracy ±0.01g), which is installed on the frame of the feeding mechanism and vibrating feeder. The sampling frequency is 200Hz, and the output is a real-time vibration acceleration value, used to detect abnormal vibrations during equipment operation.

[0076] S2. Based on the multi-source heterogeneous data, drive a dynamic digital twin model to perform online updates and calibrations to map and output the process status of the filling process in real time; remove spectral noise through wavelet transform and extract the element content corresponding to the intensity of characteristic peaks;

[0077] Furthermore, in the above technical solution, in step S2, the dynamic digital twin model is a lightweight coupling model, and its construction and operation include:

[0078] A mechanistic model framework based on material balance and reaction kinetics was constructed to simulate the physical laws governing the loading process and the molten pool response.

[0079] A real-time calibration module based on a variational autoencoder is introduced, which takes the multi-source heterogeneous data as input, dynamically infers and compensates for the system errors of the mechanistic model skeleton, and realizes online calibration of the model output.

[0080] It should be further explained that the material balance equation is: min = mout + mloss, where min is the set raw material input mass, mout is the molten pool output mass, mloss is the element burn-off mass, and mloss = min × η, where η is the element burn-off rate, which changes dynamically with temperature and reaction time.

[0081] The reaction kinetic equation is: v = k0 × e -（Ea / RT） ×c n In the formula: k0 is the pre-exponential factor, Ea is the activation energy of the reaction, R is the gas constant, T is the absolute temperature of the molten pool, c is the concentration of reactants, and n is the reaction order (calibrated to 1.2 through experiments).

[0082] Furthermore, in the above technical solution, the system error includes dynamic changes in element burn-off rate, fluctuations in heat loss, and real-time deviations in raw material reaction efficiency; the variational autoencoder encodes the latent feature distribution of the multi-source heterogeneous data and fuses it with the mechanistic model prediction in the latent space to output a calibrated state prediction.

[0083] It should be further noted that the variational autoencoder adopts an encoder-decoder structure, with the following specific configuration:

[0084] The encoder is a 3-layer fully connected network. The input layer consists of normalized multi-source data with dimension D=25+3+5=33, corresponding to 25 temperature measurement points, 3 spectral characteristic peak intensities, and 5 equipment operating parameters (feeding mechanism speed, gate opening, real-time material weight, vibration acceleration, and feeding mechanism current). The hidden layer dimensions are 256 and 128 respectively, and the final output is a latent space vector with dimension 64.

[0085] Decoder: Also a 3-layer fully connected network, the input is the concatenation of the latent vector and the predicted value of the mechanism model, and the output is the system error compensation amount, whose dimension is consistent with the output of the mechanism model (such as melt pool temperature, key components, liquid level, etc.).

[0086] Training parameters: Adam optimizer was used, learning rate was set to 0.001, number of iterations was 100, and batch size was 32.

[0087] The online calibration process is as follows:

[0088] Every 10 seconds, a multi-source data time window (containing data from the most recent 5 periods) is extracted, normalized, and then input into a variational autoencoder. The encoder encodes the latent features of the input data and performs weighted fusion with the predicted values ​​of the mechanism model in the latent space. The weight ω is dynamically adjusted according to the data credibility (credibility = data integrity × inverse of error variance). The decoder outputs the error compensation amount to correct the prediction results of the mechanism model, thus obtaining the calibrated process state (including melt pool temperature, composition, liquid level, etc.).

[0089] The model is updated every 10 seconds, employing incremental learning. The variational autoencoder network weights are fine-tuned only based on the latest 200 sets of data to avoid catastrophic forgetting. The specific implementation of system error compensation includes:

[0090] The element burn-off rate η is calculated by comparing the composition obtained from the spectral signal with the composition of the raw material fed into the furnace.

[0091] Heat loss fluctuations are corrected by comparing measured temperature field data with theoretical heat dissipation models based on Fourier's law;

[0092] The deviation in raw material reaction efficiency is obtained by back-calculation using real-time weighing data and the material balance equation.

[0093] S3. Based on the real-time process status output by the dynamic digital twin model, a hybrid intelligent drive algorithm is used for rolling optimization to generate the optimal alloy loading rate curve for the future.

[0094] Furthermore, in the above technical solution, step S3, the execution of the hybrid intelligent driving algorithm includes:

[0095] Using the dynamic digital twin model as the basic predictor, trend predictions that conform to physical laws are provided;

[0096] A spatiotemporal attention long short-term memory network is used as an intelligent compensator to learn the nonlinear relationship between multi-source data sequences in historical data and the prediction error of the basic predictor, and to perform dynamic error compensation for the current prediction result.

[0097] A model predictive control algorithm is adopted, with cost, component hit rate and process stability as multiple objectives. The compensated prediction results are rolled to optimize and solve, and the optimal alloy loading rate curve is generated.

[0098] It should be further explained that the dynamic digital twin model prediction is based on the calibrated process state. The model outputs the trend prediction of key parameters of the molten pool in the next minute, including temperature, element content such as Si and Cr, liquid level, etc., with a prediction step size of 0.5 seconds.

[0099] The input layer of the spatiotemporal attention long short-term memory network consists of a multi-source data sequence (33×50 dimensions) spanning 50 historical cycles (25 seconds). The spatiotemporal attention mechanism is set with a time window of 10 and a spatial dimension of 3 (corresponding to three types of data: raw materials, process, and equipment, respectively). The hidden layer has a dimension of 256, and the output layer is the error compensation value.

[0100] We employ a combination of offline training and online fine-tuning. Offline training uses production data from the past three months (approximately 100,000 sets), while the online phase fine-tunes the attention weights hourly using the latest 500 sets of data.

[0101] Furthermore, in the above technical solution, the process stability in the optimization objective function of the model predictive control algorithm is quantified and constrained by the second derivative of the loading rate curve to minimize the acceleration change of the actuator and ensure smooth loading action.

[0102] It should be further explained that the optimization objective function is in the form of a weighted multi-objective:

[0103] J = ω1 × J1 + ω2 × J2 + ω3 × J3; where: J1 is the cost target: J1 = α × raw material consumption + β × energy consumption (α = 0.6, β = 0.4, unit is yuan).

[0104] J2 is the component hit rate target: In the formula, c is the prediction component, c0 is the target value, Δc is the tolerance, and P is the probability of exceeding the tolerance.

[0105] J3 is the process stationarity objective: In the formula The second derivative of the loading rate is given, and the integration time is T = 10 seconds.

[0106] The weights are set to ω1=0.2, ω2=0.6, and ω3=0.2.

[0107] Optimization constraints include:

[0108] The loading rate v∈[0.5,5]kg / s;

[0109] Actuator acceleration .

[0110] The solution uses the interior point method, with a rolling time domain length of 20 control cycles (10 seconds). It is resolved every 0.5 seconds to output the optimal loading rate curve for the next 10 seconds.

[0111] S4. Based on the optimal alloy loading rate curve, a collaborative control strategy based on virtual spindle and electronic cam mapping is used to generate synchronous control commands for at least two actuators to complete precise continuous loading.

[0112] Furthermore, in the above technical solution, in step S4, the collaborative control strategy based on virtual spindle and electronic cam mapping is as follows: Figure 2 As shown, the specific implementation steps include:

[0113] S401: Integrate the optimal alloy loading rate curve into a virtual spindle displacement curve based on time or process progress.

[0114] It should be further explained that the virtual spindle displacement

[0115] Where s0 is the initial displacement (set to 0). The virtual spindle speed is set to a uniform speed ω0 = 10 rad / s, and the displacement unit is uniformly set to radians, using a scaling factor. Transformation: s(t) = k × ∫v(τ)dτ;

[0116] S402: Predefine the cam curve mapping relationship between each actuator and the virtual spindle displacement curve;

[0117] It should be further explained that, taking the three actuators of "feed belt + vibrating feeder + gate valve" as an example, the cam curve mapping relationship of the actuators is defined as follows:

[0118] Feed belt: The cam curve is a "displacement-speed" mapping, fitted using a cubic polynomial.

[0119] n(s) = a1 × s 3 +b1×s 2 +c1×s+d1, the coefficients were calibrated experimentally (a1=0.002, b1=-0.03, c1=0.5, d1=10).

[0120] Vibrating feeder: The cam curve is a "displacement-vibration frequency" mapping.

[0121] f(s) = a² × s 2 +b2×s+c2 (coefficients: a2=0.05, b2=0.3, c2=50).

[0122] Gate valve: The cam curve is a linear mapping of "displacement-opening degree".

[0123] o(s) = a3 × s + b3 (Example coefficients: a3 = 0.02, b3 = 5), ensuring that the opening degree matches the loading rate.

[0124] S403: During the control cycle, each actuator calculates and executes its own motion control command synchronously with the real-time displacement of the virtual spindle according to its own cam curve mapping relationship.

[0125] It should be further noted that the control cycle for synchronous control of multiple actuators is 0.1 seconds. Each actuator is equipped with a servo controller (position control mode, positioning accuracy ±0.001 rad).

[0126] Within each control cycle, the PLC reads the real-time displacement s(t) of the virtual spindle, calculates the corresponding target control values ​​(speed, frequency, opening degree) based on the cam curve mapping relationship of each actuator, and synchronously sends them to each servo controller via the EtherCAT bus.

[0127] During the synchronization process, feedback values ​​from each mechanism are collected in real time. If the synchronization error exceeds ±0.01 rad, the target value is corrected in the next cycle through a feedforward compensation algorithm.

[0128] Furthermore, in the above technical solution, the method also includes an adaptive replay and learning step:

[0129] Record the virtual spindle displacement curve and the actual control parameters of each actuator during the successful production process to form a standard process package;

[0130] When replaying the standard process package, the actual feedback value and command value of each actuator are compared in real time, and instantaneous fine-tuning is performed through a feedforward compensation algorithm. The basis for the fine-tuning includes the real-time deviation of the weighing sensor and the abnormal signal of the vibration sensor.

[0131] It should be further explained that the adaptive replay and learning process is as follows: Figure 3 As shown, the specific implementation process is as follows:

[0132] (a) Construction of standard process packages

[0133] The storage format is a JSON file, and the content includes:

[0134] Virtual spindle displacement curve (sampling interval 0.1 seconds);

[0135] The sequence of control parameters for each actuator;

[0136] Production condition information (raw material batch, product model, target ingredients);

[0137] The quality assessment results for this batch of production.

[0138] Screening rules: Only production batches with a comprehensive quality score of ≥90 can be used as standard process packages and stored on the local server according to product model.

[0139] (ii) Process package replay and real-time fine-tuning

[0140] Replay trigger condition: When producing the same model of product, the corresponding standard process package is automatically called to initialize the parameters of each actuator.

[0141] Real-time comparison and compensation: The actual feedback value of each actuator is collected every 0.05 seconds, and the deviation is calculated between it and the command value. .

[0142] The feedforward compensation amount is calculated using a PID controller.

[0143] Examples of compensation parameters for each institution are as follows:

[0144] Feeding mechanism: K p =0.5,K i =0.1,K d =0.05;

[0145] Vibrating feeder: K p =0.4,K i =0.08,K d =0.03;

[0146] Gate valve: K p =0.6,K i =0.12,K d =0.06.

[0147] Anomaly Handling: When the vibration sensor detects an acceleration > 0.5g, temporarily disable mechanism K. d Adjust it to twice the original value to quickly suppress vibration.

[0148] Furthermore, in the above technical solution, the method also includes a real-time filling quality assessment step:

[0149] Based on the real-time mapping of the process status in step S2, especially the dynamic deviation between the predicted and target values ​​of key components in the molten pool, the overall quality score of the current filling process is calculated.

[0150] When the overall quality score is lower than a preset threshold, the hybrid intelligent drive algorithm is triggered to start the re-optimization process and correct the subsequent loading rate curve.

[0151] It should be further explained that the real-time filling quality assessment steps are as follows: Figure 4 As shown, the specific implementation process is as follows:

[0152] (a) Calculation of overall quality score

[0153] A weighted scoring method is used, with a total score of 100 points:

[0154] Q=0.6×Q1+0.2×Q2+0.1×Q3+0.1×Q4

[0155] Where: Q1 is the component deviation score: the relative deviation is calculated for key elements (such as Si, Mn). .

[0156] Scoring rules: δ c ≤0.3 gets 100 points, 0.3<δ c ≤0.6 gets 80 points, 0.6<δ c ≤0.9 scores 60 points, δ c A score of >0.9 results in 0 points; the average score of all key elements is taken as Q1.

[0157] Q2 is the temperature deviation score: Calculation

[0158] (T0 is the target temperature ΔT = 50℃), and the scoring rules are the same as Q1.

[0159] Q3 is the process stability score: calculate the maximum value of the second derivative of the loading rate. ,like ≤0.1kg / s 3 100 points are awarded, and 20 points are deducted for every 0.1 points gained.

[0160] Q4 is the equipment operating status score: 100 points for no abnormal signals; 20 points are deducted for each minor abnormality (such as momentary vibration); 0 points are awarded for a serious abnormality (such as overload).

[0161] (II) Further Optimize Triggering and Execution

[0162] The re-optimization process is triggered when the preset quality threshold Q is less than 85 points.

[0163] Further optimization strategy adjustments include:

[0164] The MPC rolling time domain length is shortened to 10 cycles (5 seconds).

[0165] Increase the optimization frequency to once every 0.25 seconds;

[0166] Increase the target weight of component hit rate to ω2=0.7.

[0167] Correction range limit: The adjustment amount of the subsequent loading rate curve shall not exceed 5% of the original rate, so as to avoid excessive action of the actuator.

[0168] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for continuous loading control of alloy billets based on multi-source data fusion, characterized in that, Specifically, the following steps are included: S1. Real-time acquisition of multi-source heterogeneous data during the continuous loading process of alloy billets; S2. Based on the multi-source heterogeneous data, drive a dynamic digital twin model to perform online updates and calibrations, so as to map and output the process status of the filling process in real time. S3. Based on the real-time process status output by the dynamic digital twin model, a hybrid intelligent drive algorithm is used for rolling optimization to generate the optimal alloy loading rate curve for the future. S4. Based on the optimal alloy loading rate curve, a collaborative control strategy based on virtual spindle and electronic cam mapping is used to generate synchronous control commands for at least two actuators to complete precise continuous loading.

2. The method for continuous loading control of alloy billets based on multi-source data fusion according to claim 1, characterized in that, In step S1, the multi-source heterogeneous data includes: Raw material property data: composition, moisture content, and particle size distribution of the alloy material; Process status data: temperature field of melting chamber, melt surface image and intermittent component spectrum signal; Equipment operating data: feeding mechanism speed, gate opening and weighing sensor readings.

3. The method for continuous loading control of alloy billets based on multi-source data fusion according to claim 2, characterized in that, In step S1, the real-time acquisition of the process status data specifically includes: Multi-point temperature field data of the melting chamber were collected using an infrared thermal imager; The compositional spectral signals of elements on the surface of the molten pool were intermittently acquired using a laser-induced breakdown spectrometer. Visual image data of the melt surface is acquired using a high-speed industrial camera to analyze molten pool fluctuations and slag phase state.

4. The method for continuous loading control of alloy billets based on multi-source data fusion according to claim 1, characterized in that, In step S2, the dynamic digital twin model is a lightweight coupled model, and its construction and operation include: A mechanistic model framework based on material balance and reaction kinetics was constructed to simulate the physical laws governing the loading process and the molten pool response. A real-time calibration module based on variational autoencoder is introduced, which takes the multi-source heterogeneous data as input, dynamically infers and compensates for the system errors of the mechanistic model skeleton, and realizes online calibration of the model output.

5. The method for continuous loading control of alloy billets based on multi-source data fusion according to claim 4, characterized in that, The system errors include dynamic changes in element burn-off rate, fluctuations in heat loss, and real-time deviations in raw material reaction efficiency; the variational autoencoder encodes the latent feature distribution of the multi-source heterogeneous data and fuses it with the mechanistic model prediction in the latent space to output a calibrated state prediction.

6. The method for continuous loading control of alloy billets based on multi-source data fusion according to claim 1, characterized in that, In step S3, the execution of the hybrid intelligent driving algorithm includes: Using the dynamic digital twin model as the basic predictor, trend predictions that conform to physical laws are provided; A spatiotemporal attention long short-term memory network is used as an intelligent compensator to learn the nonlinear relationship between multi-source data sequences in historical data and the prediction error of the basic predictor, and to perform dynamic error compensation for the current prediction result. A model predictive control algorithm is adopted, with cost, component hit rate and process stability as multiple objectives. The compensated prediction results are rolled to optimize and solve, and the optimal alloy loading rate curve is generated.

7. The method for continuous loading control of alloy billets based on multi-source data fusion according to claim 6, characterized in that, In the optimization objective function of the model predictive control algorithm, the process stability is quantified and constrained by the second derivative of the loading rate curve to minimize the acceleration change of the actuator and ensure smooth loading action.

8. The method for continuous loading control of alloy billets based on multi-source data fusion according to claim 1, characterized in that, In step S4, the specific implementation steps of the collaborative control strategy based on virtual spindle and electronic cam mapping include: S401: Integrate the optimal alloy loading rate curve into a virtual spindle displacement curve based on time or process progress. S402: Predefine the cam curve mapping relationship between each actuator and the virtual spindle displacement curve; S403: During the control cycle, each actuator calculates and executes its own motion control command synchronously with the real-time displacement of the virtual spindle according to its own cam curve mapping relationship.

9. The method for continuous loading control of alloy billets based on multi-source data fusion according to claim 8, characterized in that, The method also includes an adaptive replay and learning step: Record the virtual spindle displacement curve and the actual control parameters of each actuator during the successful production process to form a standard process package; When replaying the standard process package, the actual feedback value and command value of each actuator are compared in real time, and instantaneous fine-tuning is performed through a feedforward compensation algorithm. The fine-tuning is based on the real-time deviation of the weighing sensor and the abnormal signal of the vibration sensor.

10. The method for continuous loading control of alloy billets based on multi-source data fusion according to claim 1, characterized in that, The method also includes a real-time filling quality assessment step: Based on the real-time mapping of the process status in step S2, especially the dynamic deviation between the predicted and target values ​​of key components in the molten pool, the overall quality score of the current filling process is calculated. When the overall quality score is lower than a preset threshold, the hybrid intelligent drive algorithm is triggered to start the re-optimization process and correct the subsequent loading rate curve.

Images