Rotating electrode atomization pulverization control system and control method

Abstract

The invention provides a rotating electrode atomization powder preparation control system and a control method. A multivariable machine learning prediction model among the electrode rotating speed, the plasma arc power, the protective gas flow, the powder particle size and the sphericity degree is established; calculating the optimal electrode rotating speed, plasma arc power and protective gas flow by using the multivariable machine learning prediction model according to the target powder particle size and sphericity; the optimal technological parameters are applied to the powder making process, and a powder making system is operated; the electrode rotating speed, the electric arc power and the gas pressure are monitored in real time in the atomization process, and technological parameters are dynamically adjusted to compensate fluctuation; the image acquisition unit is used for acquiring images of flying powder particles, particle size distribution and sphericity are analyzed in real time through an AI visual model, and the electrode rotating speed, plasma arc power and protective gas flow are fed back and adjusted according to analysis results.

Description

TECHNICAL FIELD

[0001] The application provides a rotating electrode atomization powder production control system and a control method, and relates to the technical field of atomization powder production control. BACKGROUND

[0002] The rotating electrode atomization powder production technology is an advanced process for preparing metal or alloy spherical powder. The technology generally includes a high-speed rotating metal electrode as raw material, a plasma arc heat source for melting the electrode tip, and an inert gas protection environment.

[0003] Under the action of centrifugal force generated by high-speed rotation and high temperature of the plasma arc, the electrode tip melts to form a liquid film, which is further thrown out and broken into fine droplets. The droplets spheroidize and solidify due to surface tension during flight, and finally form spherical powder. The particle size distribution and sphericity of the powder are mainly controlled by several key process parameters, including: Electrode rotation speed: directly affects the size of centrifugal force, and is a key factor for controlling the particle size of the powder; plasma arc power: determines the melting rate and superheat, and affects the fluidity and spheroidization effect of the droplets; protective gas flow and pressure: affect the atomization environment and cooling rate, and help prevent powder oxidation.

[0004] The traditional control method mainly relies on the experience of operators or simple linear models to set the rotation speed, power and gas flow. However, there is a complex nonlinear coupling relationship between these parameters, which jointly affects the particle size and sphericity of the final product. There is a lack of a mathematical model that can accurately describe the complex relationship between multiple process parameter inputs and multiple quality index outputs, resulting in a long process development cycle, high trial and error cost, and difficulty in quickly finding the optimal parameter combination for producing a specific specification of powder.

[0005] Moreover, the traditional quality detection method is offline, and the particle size and sphericity are detected by sampling, sieving and microscope observation after the entire batch of powder production is completed. This method has serious hysteresis. When the product quality is found to be unqualified, the entire batch of product may have become waste, causing huge material and energy waste. The product quality cannot be monitored in real time during the production process, let alone dynamically adjusting the process parameters according to real-time quality data.

[0006] In addition, the traditional process has poor stability and weak ability to resist external disturbances. During production, external interference factors such as electrode wear, power fluctuations and gas pressure changes are difficult to avoid. The traditional control method cannot comprehensively cope with these fluctuations, causing the process parameters to deviate from the set values, thereby causing fluctuations and inconsistencies in product quality. There is a lack of an intelligent control system that can sense parameter fluctuations in real time and automatically perform dynamic compensation. SUMMARY

[0007] In view of the above technical problems, the present application provides a rotating electrode atomization powder production control method, comprising the following steps: establishing a multivariate machine learning prediction model between the electrode rotating speed, the plasma arc power and the protective gas flow and the particle size and sphericity of the powder; According to the particle size and sphericity of the target powder, the optimal process parameters are calculated using the multivariate machine learning prediction model, including the optimal electrode rotating speed, the optimal plasma arc power and the optimal protective gas flow; The optimal process parameters are applied to the powder production control system; In the rotating electrode atomization powder production process, the process parameters are monitored and dynamically adjusted in real time to maintain them within the fluctuation range of the optimal process parameters; The image acquisition unit is used to obtain the image of the powder particles in flight, the AI vision model is used to analyze the particle size and sphericity in real time, and the setting of the optimal electrode rotating speed, the optimal plasma arc power and the optimal protective gas flow is adjusted according to the analysis results.

[0008] Preferably, the input parameters of the multivariate machine learning prediction model are fusion features including original inputs and physical constraint inputs, The original inputs are: electrode rotating speed, plasma arc power, protective gas flow; The physical constraint inputs are: centrifugal term and particle size constraint, sphericity and power and flow constraint; The original inputs and physical constraint inputs are spliced into fusion features as input parameters of the multivariate machine learning prediction model.

[0009] Preferably, the centrifugal term and particle size constraint are: ; Wherein, k1 is an empirical coefficient, ω is the electrode rotating speed, r is the centrifugal radius, g is the acceleration of gravity, Indicates the particle size of the powder; The sphericity and power and flow constraint are: ; Wherein, S is the sphericity, P is the plasma arc power, Q is the protective gas flow, Is the adjustment parameter.

[0010] Preferably, the step of analyzing the particle size and sphericity in real time through the AI vision model comprises: (1) removing the noise of the image of the powder particles by using Gaussian filtering or median filtering, separating the powder particles from the background by semantic segmentation, and then extracting single particles and assigning unique identification through contour detection; (2) predicting the particle size and sphericity from the image of the powder particles by using a regression neural network; (3) updating the particle size histogram and sphericity qualification rate in real time.

[0011] Preferably, the step of adjusting the settings of the optimal electrode rotation speed, the optimal plasma arc power and the optimal shielding gas flow according to the analysis result comprises: increasing the optimal electrode rotation speed when the average particle size is continuously detected to increase; increasing the optimal plasma arc power when the average particle size is continuously detected to be less than the target particle size range; and adjusting the optimal shielding gas flow when the satellite powder increases to cause the sphericity to decrease.

[0012] Preferably, when the particle size D 50 When the deviation exceeds ± 5% or the sphericity qualification rate is lower than 95%, an alarm is triggered and an adjustment instruction is issued to the control system through the PLC.

[0013] The application further provides a rotating electrode atomization powder production control system for realizing the rotating electrode atomization powder production control method, which is characterized by comprising a sensing system, a data preprocessing module, a model operation module, an execution control module and a quality detection module, and each module realizes data interaction through an industrial Ethernet. The sensing system is a distributed sensor network, which comprises a rotation speed sensor installed at the end of an electrode driving shaft, a power sensor integrated in a plasma arc generator and a gas flow sensor arranged in a shielding gas pipeline, and all sensor data is collected to the data preprocessing module through the industrial Ethernet. The data preprocessing module sequentially performs outlier rejection and standardization on the original data and outputs the processed data to the model operation module. The model operation module divides the received data into two paths: one path is input into a multivariate machine learning model, and the other path is stored in a historical database to support incremental learning; the model operation module can realize forward prediction and reverse optimization, and realizes incremental learning by calling the latest data in the historical database, and the operation result is transmitted to the execution control module. The execution control module is composed of a PLC controller and an execution mechanism, the PLC controller receives the instruction, adjusts the electrode rotation speed through a servo driver, adjusts the plasma arc power through a power regulator and adjusts the shielding gas flow through a proportional valve, and when the process parameters fluctuate beyond the fluctuation range, a compensation algorithm is automatically started to correct the process parameters. The quality detection module comprises a high-speed camera and an AI vision processor, the high-speed camera photographs powder particles, the AI processor calculates the particle size and sphericity through semantic segmentation and feeds back to the model operation module, and the multivariate machine learning model is triggered to optimize the learning parameters.

[0014] Preferably, in the model operation module, when the single prediction error of the particle size exceeds ± 4.1 mu m or the single prediction error of the sphericity exceeds ± 0.015 during forward prediction, the historical data with a similarity of greater than or equal to 90% in the historical database are automatically called to fine-tune the learning parameters of the multivariate machine learning model.

[0015] Preferably, in the execution control module, the servo driver adopts permanent magnet synchronous motor drive, the electrode rotating speed regulation precision is ±5rpm, the power regulator is provided with a voltage compensation module, the influence of power grid voltage ±10% fluctuation on arc power is offset in real time, and the compensation response time is less than 50ms; the proportional valve adopts electro-hydraulic proportional control mode, and the protection gas flow control precision is ±2SLM.

[0016] Compared with the prior art, the present application has the following beneficial technical effects: 1. The present application directly links the electrode rotating speed, arc power, gas flow and other core process parameters with the particle size and sphericity of the powder, which are two key quality indicators, by establishing a multivariate machine learning prediction model. The system can not only deduce the optimal process parameters according to the target product quality, but also dynamically feedback and adjust according to real-time monitoring data during production, thereby significantly improving the control accuracy of the final product particle size distribution and sphericity, and ensuring the stability and consistency of the powder product quality.

[0017] 2. The present application inputs the centrifugal term, particle size constraint and sphericity and power flow constraint as feature inputs into the machine learning model, fuses the original data with physical constraints, enhances the physical meaning of the model, and makes the decision not only based on data statistics, but also more in line with the physical mechanism of the atomization powder production process, which helps the model to converge faster, improves the prediction reliability in unknown process intervals, and reduces the excessive dependence on a large amount of pure data training.

[0018] 3. A high-efficiency real-time closed-loop feedback control loop is constructed: the system realizes online, non-destructive and real-time detection of flying powder particles through high-speed cameras and AI vision models. Traditional methods often need to interrupt production for offline screening and microscope observation, which has serious lag. The present application can immediately obtain particle size distribution and sphericity information through real-time image analysis, and quickly feedback the results to the control system to timely suppress production fluctuations and reduce the scrap rate. BRIEF DESCRIPTION OF DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0020] Figure 1 Flow chart of the rotating electrode atomization powder control method of the present application; Figure 2 The outline map of each particle is separated by semantic segmentation; Figure 3 Metal droplet image; Figure 4 Fig. 1 is a structural schematic diagram of a model operation module, an execution control module and a quality detection module. DETAILED DESCRIPTION

[0021] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0022] In the drawings of the specific embodiments of the present application, in order to better and more clearly describe the working principles of the elements in the system and show the connection relationship of the parts in the device, only the relative position relationship between the elements is distinguished, and the signal transmission direction, connection order and the size and shape of the parts in the elements or structure cannot be limited.

[0023] Embodiment 1 As shown in Fig. 1, it is a flowchart of the control method of the rotating electrode atomization powder production of the present application, including the following steps: Figure 1 I. Establishing a multivariate machine learning prediction model between the electrode rotating speed, the plasma arc power and the protective gas flow and the particle size and sphericity of the powder. I. Establishing a multivariate machine learning prediction model between the electrode rotating speed, the plasma arc power and the protective gas flow and the particle size and sphericity of the powder.

[0024] 1. Data preparation and preprocessing In order to meet the precise correlation requirements of the process parameters and product performance indicators in the metal powder production process, data preparation and preprocessing are first carried out.

[0025] The collected data includes: the electrode rotating speed ω, the plasma arc power P and the protective gas flow Q.

[0026] In the metal powder centrifugal atomization production process, the electrode rotating speed, the plasma arc power and the protective gas flow are the core input variables that determine the powder quality, among which the electrode rotating speed directly affects the centrifugal force suffered by the metal droplet. The higher the rotating speed, the more sufficient the droplet breaking, and the smaller the particle size.

[0027] The numerical range of the electrode rotation speed ω is 2000-15000 rpm, covering different preparation requirements from coarse powder to fine powder; the plasma arc power determines the melting degree and temperature of the metal raw material, and too low power will lead to incomplete melting of the metal, forming irregular particles, and too high power will easily cause excessive evaporation of the metal, so the numerical range of the plasma arc power is determined as 80-200 kW; the protective gas flow Q is used to isolate air, prevent metal oxidation and assist droplet cooling, and insufficient protective gas flow will lead to increased oxygen content, and too high flow will blow off the un-solidified droplets, so the numerical range of the protective gas flow Q is 30-100 SLM.

[0028] In terms of target output variables, particle size D 50 As a key indicator representing the particle size distribution of the powder, it directly affects the flowability and formability of the powder, and the particle size D 50 The target range is 15-150 μm, which can adapt to different requirements of additive manufacturing process; the sphericity reflects the regularity of the powder particles, and high sphericity powder can improve the density of the formed part, and the sphericity needs to be controlled between 0.85-0.99.

[0029] To ensure the generalization ability and reliability of the prediction model, the data set needs to cover at least 500 experimental records, and cover Ti6Al4V, 316 stainless steel and other commonly used metal materials, among which Ti6Al4V has excellent high temperature strength and corrosion resistance, and 316 stainless steel has good biocompatibility and corrosion resistance.

[0030] 2. Data cleaning and standardization During the experiment, due to factors such as equipment running stability, measurement instrument accuracy and operating environment, there may be extreme values and noise in the data set, such as instantaneous out-of-range fluctuations of the electrode rotation speed due to equipment failure, abnormal peak values of the plasma arc power due to unstable voltage, or abnormal data of the protective gas flow due to valve control error. If these data are directly used for modeling, it will deviate seriously from the true process rule, leading to a decrease in the prediction accuracy of the prediction model. Therefore, statistical methods are needed to filter the data, remove extreme values and noise outside the reasonable range, and normalize the input and output variables to the range of [0, 1]. Normalization can eliminate the influence of dimension, so that all variables are in the same order of magnitude, ensuring that the prediction model can learn the influence rule of each variable uniformly during gradient descent, improving training efficiency and stability of the prediction model.

[0031] The processed data set is randomly divided into a training set, a validation set and a test set in a ratio of 6:2:2, wherein the training set accounts for 60%, is used for learning and fitting of prediction model parameters, and needs to contain a sufficient number of process parameter combinations to ensure that the prediction model can fully learn the correlation between the input and the output; the validation set accounts for 20%, is mainly used for adjustment and optimization of the learning parameters of the prediction model, and the performance indicators of the validation set are monitored to avoid overfitting of the prediction model; the test set accounts for 20%, needs to be independent of the training set and the validation set, is used to simulate real application scenarios, and objectively evaluates the generalization ability and prediction accuracy of the prediction model, and in the division process, a random sampling method is used, and the input variable distribution and the output variable distribution of each data set are ensured to be consistent with the original data set, so as to avoid distortion of the evaluation results due to data distribution deviation.

[0032] 3. Obtain physical constraint relationship combined with univariate modeling To further improve the physical rationality and prediction accuracy of the prediction model, univariate modeling and physical relationship calibration need to be carried out to establish the basic physical correlation between the process parameters and the output indicators, and to provide constraint basis for subsequent multivariate prediction models.

[0033] The physical constraint relationship combined with univariate modeling is obtained, and two physical constraint features are constructed: the first constraint feature is the correlation index of the centrifugal term and the particle size, and the empirical coefficient in the univariate model is used to convert the electrode rotating speed into the centrifugal term feature directly related to the particle size; the second constraint feature is the correlation index of the sphericity and the power and the flow, and the adjustment parameter in the sphericity model is used to fuse the plasma arc power and the protective gas flow into a comprehensive feature for representing the synergistic effect of the two on the sphericity.

[0034] (1) Modeling of the physical constraint of the centrifugal term and the particle size Particle size

[0035] wherein, is the particle size, ω is the electrode rotating speed, r is the centrifugal radius, g is the gravitational acceleration, and k1 is the empirical coefficient.

[0036] In the process of centrifugal atomization for preparing metal powder, the breaking and particle size formation of metal droplets are mainly driven by centrifugal force. According to the physical principle, the centrifugal force is proportional to the square of the electrode rotating speed and the electrode radius, and is inversely proportional to the gravitational acceleration.

[0037] Let , the empirical coefficient k1 is solved by linear regression fitting y=k1x.

[0038] Preferably, 100 sets of experimental data of Ti6Al4V alloy are selected for fitting, the physical properties of Ti6Al4V such as melting point and density are stable, the repeatability and reliability of the experimental data are high, the least square method is used in the fitting process, the optimal empirical coefficient k1 is determined by minimizing the residual sum of squares of the actual particle size and the predicted particle size, the empirical coefficient can quantitatively reflect the influence degree of the centrifugal term on the particle size of Ti6Al4V alloy in the centrifugal atomization process, and provide important physical constraints for subsequent multivariate model prediction of particle size, avoid the prediction result that the particle size increases with the increase of the rotating speed, which violates the physical law, and finally the fitting result of .

[0039] (2) Modeling of the physical constraints of sphericity S and plasma arc power P, and protective gas flow Q:

[0040] wherein, is an adjustment parameter, and tanh is a hyperbolic tangent function.

[0041] The plasma arc power determines the melting temperature and residence time of the metal droplet, when the power is too low, the metal raw material cannot be completely melted, and the droplet cannot form a regular sphere during the cooling process; when the power is too high, the droplet is overheated, and evaporation or splashing phenomenon occurs, which also affects the sphericity. The protective gas flow affects the sphericity by controlling the cooling speed and isolating air, when the flow is insufficient, the cooling speed of the droplet is too slow, and the droplet is easily oxidized by air to form an irregular surface; when the flow is too high, the impact force of the gas flow on the droplet increases, which may cause the droplet to deform. The adjustment parameter is introduced to balance the interaction between the two. The determination of the adjustment parameter needs to be optimized through multiple sets of cross experimental data.

[0042] 3. Construction of multivariate machine learning prediction model Based on the results of the previous data processing and physical constraints, the construction of the multivariate machine learning prediction model is carried out, and by fusing the original input and the physical constraint input, the prediction accuracy of the model for the particle size D 50 and the sphericity is improved.

[0043] (1) Fusion feature construction Original input: electrode rotating speed, plasma arc power, and protective gas flow.

[0044] Physical constraint input: based on the results of the single variable model, two physical constraints are input, i.e. the centrifugal term and the particle size constraint, and the sphericity and the power and flow constraint.

[0045] Fusion feature: The original input and the physical constraint input are spliced into a fusion feature as an input parameter of the multivariate machine learning prediction model, i.e., 3 original variables and 2 physical constraints, a total of 5-dimensional input.

[0046] The 3 original variables and 2 physical constraints are spliced to form a 5-dimensional input fusion feature. The fusion feature not only retains the parameter information of the original input, but also introduces constraints based on physical laws, which can guide the prediction model to learn key influencing factors more efficiently and avoid the model from overfitting to irrelevant features.

[0047] (2) Structure of the prediction model A multilayer perceptron neural network is used as the prediction model. The multilayer perceptron has strong non-linear fitting capability and can effectively handle the complex non-linear relationship between process parameters and output indicators, making it suitable for such a complex system with multiple factors coupling as metal powder preparation.

[0048] The structure of the prediction model is as follows: 5 nodes are set in the input layer, corresponding to the 5-dimensional fusion feature, to ensure that all key information can be effectively input into the prediction model; 2 layers are set in the hidden layer, the first layer contains 128 neurons, and the second layer contains 64 neurons, with a decreasing number of neurons to gradually extract key information from the input features and reduce the interference of redundant features. The first layer captures the complex interaction between features through more neurons, and the second layer integrates and compresses the key features through fewer neurons; the activation function is ReLU function, which can effectively solve the gradient vanishing problem compared with traditional Sigmoid or Tanh function, improve the training efficiency of deep network, and has lower computational complexity, which helps to speed up the model convergence speed.

[0049] Preferably, to prevent model overfitting, a double regularization mechanism is introduced: Specifically, L2 regularization is used with a regularization coefficient of 0.0001. By adding a weight square term in the loss function, the prediction model weight is limited to avoid overfitting of the prediction model to some features; at the same time, an early stopping mechanism is used to monitor the loss value of the validation set in real time during the training of the prediction model. When the loss value of the validation set does not decrease continuously for multiple rounds, the training is stopped immediately to prevent the model from overfitting to the training set after further training.

[0050] The loss function preferably adopts mean square error, which effectively measures the deviation between the predicted value and the true value; the optimizer preferably adopts Adam optimizer, which combines the advantages of momentum gradient descent and adaptive learning rate, can adaptively adjust the learning rate of each parameter, has faster convergence speed compared with traditional stochastic gradient descent, and is not easy to fall into local optimal solution, and the learning rate is set to 0.0005, which is adjusted through multiple pre-experiments, which can ensure fast convergence of the model and avoid training shock caused by too large learning rate.

[0051] To determine the optimal prediction model learning parameter combination, learning parameter search work is carried out, the search range includes the number of hidden layers (1 layer, 2 layers, 3 layers), the number of neurons in each layer (32, 64, 128), the activation function (ReLU function and Tanh function) and the regularization coefficient (0.0001, 0.001, 0.01), and the grid search method is used to traverse all learning parameter combinations, and the mean square error of the validation set is used as the evaluation standard to select the learning parameter combination with the smallest mean square error, so as to ensure that the model has good fitting ability and excellent generalization performance.

[0052] 4. Model verification and performance evaluation The historical data set used in this evaluation contains 500 complete metal powder preparation experiment records, covering different process parameter combinations and various metal materials, and the value range of the input variables strictly follows the actual production requirements. In the data division process, random sampling combined with stratified sampling method is adopted, and the data set is divided into training set, validation set and test set according to the ratio of 6:2:2, stratified sampling can ensure that the input variable distribution and output variable distribution of each data set are consistent with the original data set, avoid the distortion of evaluation results caused by data distribution deviation, and the test set is completely independent of the data set, which is not involved in the training and learning parameter optimization process of the model, and can truly reflect the generalization ability of the model in the actual application scene.

[0053] The performance evaluation adopts two core quantitative indicators: mean absolute error and determination coefficient, which measure the model performance from two dimensions: error size and model explanation ability.

[0054] For the prediction of particle size D 50 , the average absolute error of the prediction model is controlled within ±4.1 μm, and the smaller the average absolute error, the smaller the average deviation between the predicted value and the true particle size, which can meet the precision requirements of particle size control in industrial production.

[0055] Specifically, in the preparation of Ti6Al4V powder in the aerospace field, the particle size deviation needs to be controlled within 5 μm, and the prediction model fully meets the requirements; at the same time, the determination coefficient R2 of particle size prediction is greater than 0.95, and the closer the determination coefficient is to 1, the higher the degree of particle size variation that the prediction model can explain. When R2 is greater than 0.95, it indicates that the prediction model can capture more than 95% of the particle size variation rule, and the prediction result has very high reliability. For the prediction of sphericity, the average absolute error of the prediction model is ±0.015, and sphericity is an index reflecting the regularity of particles. In industrial production, the prediction deviation is required to be less than 0.02, and the error of the prediction model is fully within the acceptable range; the determination coefficient R2 of sphericity prediction is greater than 0.89, although it is slightly lower than the R2 of particle size prediction, but considering that sphericity is affected by more subtle factors such as metal droplet cooling speed and airflow disturbance, the determination coefficient can fully illustrate the ability of the model to capture the variation rule of sphericity, and can provide accurate sphericity prediction for actual production.

[0056] Preferably, to ensure the reproducibility of the evaluation results, a 5-fold cross-validation method is used to divide the data set into 5 non-overlapping subsets, and each time 4 subsets are selected as the training set and 1 subset is selected as the test set. The average performance index is calculated by repeating the experiment 5 times, and the results show that the average absolute error and the determination coefficient of the 5 experiments fluctuate little, proving that the performance of the model is stable and reliable, and can maintain consistent prediction accuracy on different data subsets. In addition, the model is compared with the traditional linear regression model and the single variable prediction model, and the prediction R2 of the traditional model for particle size D 50 is only about 0.8, and the prediction R2 for sphericity is less than 0.75, while the R2 of the model is increased by more than 15%, fully proving the significant advantage of the multivariate machine learning model fused with physical constraints in prediction accuracy, which can provide strong support for parameter optimization and quality control of metal powder preparation process.

[0057] II. According to the particle size and sphericity of the target powder, the optimal process parameters are calculated by using the multivariate machine learning prediction model, including the optimal electrode speed, the optimal plasma arc power and the optimal protective gas flow.

[0058] Input target powder performance, in a specific embodiment, D 50 = 50 μm, sphericity S ≥ 0.95, the combination of optimal process parameters is optimized by reverse prediction, and the optimal process parameters are output: optimal electrode speed, optimal plasma arc power and optimal protective gas flow.

[0059] Combined with the real-time monitoring data of the online particle size analyzer and temperature sensor in production, the learning parameters of the prediction model are dynamically corrected, and the prediction model is incrementally trained with new experimental data to maintain consistency between the prediction model and the actual process.

[0060] III. Apply the optimal process parameters to the powder production process, and run the powder production system.

[0061] In the sealed chamber filled with inert gas, start the driving device to rotate the electrode at high speed. After the speed stabilizes, ignite the coaxial plasma arc to uniformly melt the end face of the electrode rod and form a liquid film.

[0062] At this stage, ensure that the rotation, arc, cooling, and gas protection system are running smoothly, and the process parameters are stable at the optimal values calculated in step one, to create stable conditions for the uniform formation and atomization of droplets.

[0063] Construct a dynamic compensation system. During the atomization process, use a sensor network to monitor the actual speed of the electrode, the power of the plasma arc, and the fluctuation of the protective gas flow in the chamber in real time, reduce the mutual interference between process parameters, and improve the stability of operation.

[0064] Compare the monitoring data with the optimal process parameters, which are the optimal electrode speed, optimal plasma arc power, and optimal protective gas flow calculated earlier. Through the control system, dynamically adjust the electrode speed, plasma arc power, and protective gas flow to compensate for possible fluctuations and ensure that the process is always in the best state.

[0065] IV. Use the image acquisition unit to obtain images of the powder particles in flight, analyze the particle size and sphericity in real time through the AI vision model, and adjust the settings of the optimal electrode speed, optimal plasma arc power, and optimal protective gas flow based on the analysis results.

[0066] The image acquisition unit uses real-time high-resolution imaging technology combined with an AI vision model to analyze the sphericity and particle size of the powder online, enabling rapid quality feedback and optimizing the production process.

[0067] In the preferred embodiment, the AI vision model combines image processing with a deep learning AI model, and the analysis method is as follows: The image acquisition unit continuously acquires image sequences of the powder flow and uses a filtering algorithm to reduce image noise.

[0068] Use semantic segmentation based on deep learning to separate the powder particles from the background in the image. Semantic segmentation can better handle complex situations such as particle overlap, adhesion, and uneven lighting, and accurately separate the outline of each particle, as shown in Figure 2 .

[0069] Perform contour finding on the segmented binary image to identify each independent particle region and assign a unique identifier to each particle.

[0070] For each extracted particle region, analyze the particle size distribution calculation and sphericity analysis.

[0071] The system runs continuously, analyzing thousands of particles per second. The particle size histogram and sphericity pass rate are updated in real time.

[0072] Set a quality threshold, preferably when the particle size D 50 When the deviation exceeds ± 5% or the sphericity pass rate is less than 95%, an alarm is triggered, prompting operator intervention. The ultimate goal is to achieve AI system correlation analysis of particle size quality indicators with electrode speed, plasma arc power, and protective gas flow.

[0073] Preferably, the AI vision model detects a continuous increase in average particle size, which can automatically determine and suggest increasing the electrode speed, and the speed increase reduces the particle size.

[0074] When the AI vision model detects that the average particle size is consistently smaller than the target particle size range, increase the optimal plasma arc power. Insufficient plasma arc power will result in insufficient energy input to the metal melt, and the melting amount or breaking energy during atomization will be low, resulting in a smaller overall particle size after atomization. At this time, the optimal plasma arc power should be increased to increase the energy and melting rate of the melt, allowing the atomization process to produce powder particles that meet the target size.

[0075] When the AI vision model detects an increase in satellite powder and a decrease in sphericity, it can determine and suggest adjusting the gas flow to reduce the collision and adhesion of droplets. For example, Figure 3 as shown.

[0076] These optimization suggestions can be displayed on the interface for manual confirmation and execution, or automatically issued by the PLC controller to the execution mechanism for real-time process optimization.

[0077] Based on the quality evaluation results, decide whether to adjust the parameters for the next batch of production.

[0078] Example 2 The present application also proposes a rotating electrode atomization powder production control system, which realizes precise control of the metal powder production process through a sensing system, data acquisition module, model operation module, execution control module, and quality detection module.

[0079] The sensing system acquires key process parameters and environmental data in real time through a distributed sensor network.

[0080] A rotational speed sensor is installed at the end of the electrode drive shaft to capture the rotational speed of the electrode at a sampling frequency of 1 kHz, with an accuracy of ±10 rpm. A power sensor is integrated into the plasma arc generator to monitor the power output in real time within the range of 80-200 kW, with an error controlled within ±1 kW. A gas flow sensor is deployed in the protective gas pipeline to dynamically record the flow changes within the range of 30-100 SLM, with a response time less than 1 s. In addition, an oxygen content sensor and a temperature sensor are used to monitor the purity of the inert gas and the working temperature in the chamber, respectively. All sensor data are aggregated through an industrial Ethernet to a data preprocessing module.

[0081] The data preprocessing module performs multiple steps of processing on the collected raw data: first, extreme outliers are removed by the 3σ criterion to eliminate noise caused by transient fluctuations of the equipment; then, Z-Score standardization is used to map each parameter to the [0, 1] interval, solving the problem of dimension difference; finally, the data trend features are extracted through a sliding window algorithm to provide stable input for the model operation module.

[0082] The data set received by the model operation module is divided into two paths, one of which is transmitted into a multivariate machine learning model for prediction, and the other is stored in a historical database to support subsequent model incremental learning. The model operation module realizes dual functions based on the machine learning model built on the PyTorch framework: forward prediction calculates the particle size and sphericity through 5-dimensional input, with a prediction error controlled within ±4.1 μm for single prediction error of particle size and ±0.015 for single prediction error of sphericity; reverse optimization outputs the optimal process parameter combination according to the user's set target performance, combining reinforcement learning algorithm: optimal electrode rotational speed, optimal plasma arc power, and optimal protective gas flow.

[0083] The model operation module performs incremental learning by calling the latest data from the historical database in real time, ensuring that the prediction accuracy is maintained when the equipment ages or the material is replaced. The operation results are transmitted to the execution control module through the OPC UA protocol.

[0084] The execution control module is composed of a PLC controller and an execution mechanism. After receiving the parameter instructions output by the model, the PLC adjusts the electrode rotational speed through a servo driver, stabilizes the arc output through a power regulator, and controls the gas flow through a proportional valve. The response time of all adjustment actions is less than 100 ms. When the parameter fluctuations detected by the sensing system exceed the threshold, the execution control module automatically starts the compensation algorithm to dynamically correct the execution process parameters, maintaining the process stability.

[0085] In a preferred embodiment, the quality detection module comprises an image acquisition unit, As Figure 4 shown, the image acquisition unit comprises a high-speed camera and an AI vision processor.

[0086] High-speed camera: installed on the powder flow path behind the atomization chamber, preferably in the settling section or inert gas circulation loop, with sufficient frame rate to capture high-speed flying powder particles and sufficient resolution to clearly present the micro-morphology of individual particles. The camera takes pictures of flying powder particles at a frame rate of 30 fps, and the LED backlight ensures image clarity.

[0087] The AI vision processor extracts particle contours through semantic segmentation, calculates particle size and sphericity in real time, and feeds back the detection results to the model operation module. When the detection value deviates from the target value by ±5%, trigger the model to re-optimize the parameters. The AI vision processor uses a GPU acceleration card to accelerate the inference process of the deep learning model, meeting the real-time requirements.

[0088] The quality detection module preferably also includes a high-brightness, uniform backlight or sidelight light source: provides stable and uniform illumination, ensures clear particle image contours and high contrast, reduces shadows and glare, and facilitates subsequent image processing. LED surface light sources are usually used for backlighting to form bright-field images.

[0089] The quality detection module preferably also includes a sealed optical window: an optical viewing window is opened on the atomization chamber, and a pollution-proof design is used to prevent powder from adhering and blocking the view.

[0090] The above modules interact with each other through industrial Ethernet to form a control system. The sensing system and data acquisition module provides basic data for the system, the pre-processing module ensures data quality, the model operation module provides intelligent decision-making, the execution control module realizes physical regulation, and the quality detection module ensures control effect, together realizing high-precision, intelligent control of metal powder preparation.

[0091] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto. Any skilled person in the art can easily think of various equivalent modifications or replacements within the technical scope disclosed in the present application, and these modifications or replacements should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A control method for atomizing powder production by a rotating electrode, characterized by, The method comprises the following steps: establishing a multivariate machine learning prediction model between electrode rotating speed, plasma arc power and protective gas flow and particle size and sphericity of the powder; calculating optimal process parameters, including optimal electrode rotating speed, optimal plasma arc power and optimal protective gas flow, according to particle size and sphericity of the target powder by using the multivariate machine learning prediction model; applying the optimal process parameters to the powder production control system; monitoring and dynamically adjusting the process parameters in real time during the rotating electrode atomization powder production process to maintain them within the fluctuation range of the optimal process parameters; acquiring images of the powder particles in flight by using an image acquisition unit, analyzing particle size and sphericity in real time by using an AI vision model, and adjusting the settings of the optimal electrode rotating speed, the optimal plasma arc power and the optimal protective gas flow according to the analysis results.

2. The rotary electrode atomizing powder production control method according to claim 1, characterized by, The input parameters of the multivariate machine learning prediction model are fusion features including original inputs and physical constraint inputs, the original inputs are electrode rotating speed, plasma arc power and protective gas flow; the physical constraint inputs are centrifugal terms and particle size constraints, sphericity and power and flow constraints; the original inputs and the physical constraint inputs are spliced into fusion features as the input parameters of the multivariate machine learning prediction model.

3. The rotating electrode atomization powder production control method according to claim 2, wherein The centrifugal term and the particle size constraint are: ; where k1 is an empirical coefficient, ω is the electrode rotational speed, r is the centrifugal radius, and g is the gravitational acceleration, D50 represents the particle size of the powder; Sphericity with power and flow rate constraints are: ; where S is the sphericity, P is the plasma arc power, Q is the shielding gas flow rate, is a tuning parameter, and tanh is the hyperbolic tangent function.

4. The rotary electrode atomizing powder production control method according to claim 1, characterized by, the step of analyzing particle size and sphericity in real time by using an AI vision model comprises: (1) removing noise of the image of the powder particles by using Gaussian filtering or median filtering, separating the powder particles from the background by using semantic segmentation, and extracting single particles and assigning unique identifiers by using contour detection; (2) predicting particle size and sphericity from the image of the powder particles by using a regression neural network; and (3) updating a particle size histogram and a sphericity qualification rate in real time.

5. The rotary electrode atomizing powder production control method according to claim 1, characterized by, The step of adjusting the settings of the optimal electrode rotating speed, the optimal plasma arc power and the optimal protective gas flow according to the analysis results comprises: increasing the optimal electrode rotating speed when it is detected that the average particle size continuously increases; increasing the optimal plasma arc power when it is detected that the average particle size continuously decreases to be less than the target particle size range; and adjusting the optimal protective gas flow when it is detected that the sphericity decreases due to an increase in satellite powder.

6. The rotary electrode atomizing powder production control method according to claim 5, characterized by, When the particle size D 50 When the deviation exceeds ± 5% or the sphericity qualification rate is lower than 95%, an alarm is triggered and an adjustment instruction is issued to the control system through the PLC.

7. A control system for a rotating electrode atomization powder production system for implementing the control method for a rotating electrode atomization powder production system according to any one of claims 1 to 6, characterized by The system comprises a sensing system, a data preprocessing module, a model operation module, an execution control module and a quality detection module, and the modules realize data interaction through industrial Ethernet; The sensing system is a distributed sensor network, which comprises a rotating speed sensor installed at the end of the electrode driving shaft, a power sensor integrated in the plasma arc generator and a gas flow sensor arranged in the protective gas pipeline, and all sensor data is collected to the data preprocessing module through industrial Ethernet; The data preprocessing module sequentially performs outlier rejection and standardization on the original data, and outputs the processed data to the model operation module; The model operation module divides the received data into two paths: one path is input into the multivariate machine learning model, and the other path is stored in the historical database to support incremental learning; The model operation module can perform forward prediction and reverse optimization, and realizes incremental learning by calling the latest data in the historical database, and the operation result is transmitted to the execution control module; The execution control module is composed of a PLC controller and an execution mechanism, the PLC controller receives instructions, adjusts the electrode rotating speed through a servo driver, adjusts the plasma arc power through a power regulator, and adjusts the protective gas flow through a proportional valve, when the process parameter fluctuation exceeds the fluctuation range, the compensation algorithm is automatically started to correct the process parameter; The quality detection module includes a high-speed camera and an AI vision processor, the high-speed camera photographs powder particles, the AI processor calculates the particle size and sphericity through semantic segmentation and feeds back to the model operation module to trigger the multivariate machine learning model to optimize the learning parameters.

8. The rotary electrode atomizing powder production control system according to claim 7, wherein In the model operation module, when the single prediction error of the particle size exceeds ±4.1 μm or the single prediction error of the sphericity exceeds ±0.015 during forward prediction, the historical data with a similarity of greater than or equal to 90% in the historical database are automatically called to fine-tune the learning parameters of the multivariate machine learning model.

9. The rotary electrode atomizing powder production control system according to claim 7, wherein In the execution control module, the servo driver adopts a permanent magnet synchronous motor drive, the electrode rotating speed adjustment precision is ±5 rpm, the power regulator has a built-in voltage compensation module, which can offset the influence of ±10% fluctuation of the power grid voltage on the arc power in real time, the compensation response time is less than 50 ms, and the proportional valve adopts an electro-hydraulic proportional control mode, and the protective gas flow control precision is ±2 SLM.

Images