A rotary electrode atomization powder production control system and control method

Abstract

The application provides a rotating electrode atomization powder production control system and a control method, a multivariate machine learning prediction model is established between an electrode rotating speed, a plasma arc power, a protective gas flow and powder particle size and sphericity, according to target powder particle size and sphericity, the multivariate machine learning prediction model is used to calculate optimal electrode rotating speed, plasma arc power and protective gas flow, the optimal process parameters are applied to a powder production process, and a powder production system is operated, the electrode rotating speed, the arc power and the gas pressure are monitored in real time in the atomization process, and the process parameters are dynamically adjusted to compensate for fluctuations, an image acquisition unit is used to obtain images of powder particles in flight, AI vision model is used to analyze particle size distribution and sphericity in real time, and the electrode rotating speed, the plasma arc power and the protective gas flow are adjusted according to the analysis results.

Description

Technical Field

[0001] This invention proposes a rotating electrode atomization powder production control system and control method, which relates to the field of atomization powder production control technology. Background Technology

[0002] Rotating electrode atomization powder preparation technology is an advanced process for preparing spherical powders of metals or alloys. This technology typically includes a high-speed rotating metal electrode as the raw material, a plasma arc heat source for the molten electrode tip, and an inert gas protective environment.

[0003] Under the centrifugal force generated by high-speed rotation and the high temperature of the plasma arc, the electrode tip melts to form a liquid film. This liquid film is further ejected and broken into tiny droplets. During flight, the droplets spherize due to surface tension and then cool and solidify, ultimately forming spherical powder. The powder particle size distribution and sphericity in this process are mainly controlled by several key process parameters, including:

[0004] Electrode rotation speed: directly affects the magnitude of centrifugal force and is a key factor in controlling powder particle size; Plasma arc power: determines the melting rate and superheat, affecting droplet flowability and spheroidization effect; Protective gas flow rate and pressure: affect the atomization environment and cooling rate, and help prevent powder oxidation.

[0005] Traditional control methods mainly rely on operator experience or simple linear models to set rotational speed, power, and gas flow rate. However, these parameters have complex nonlinear coupling relationships that jointly affect the particle size and sphericity of the final product. The lack of a mathematical model that can accurately describe the complex relationship between multiple process parameter inputs and multiple quality index outputs leads to long process development cycles, high trial-and-error costs, and difficulty in quickly finding the optimal parameter combination for producing powders of specific specifications.

[0006] Furthermore, traditional quality inspection methods are offline, involving sampling, sieving, and microscopic observation to test particle size and sphericity after the entire batch of powder has been produced. This method suffers from significant time lag. By the time product quality is found to be substandard, the entire batch may already be scrap, resulting in enormous waste of materials and energy. Real-time monitoring of product quality is impossible during production, let alone dynamically adjusting process parameters based on real-time quality data.

[0007] Furthermore, traditional processes suffer from poor stability and weak resistance to external disturbances. During production, external interference factors such as electrode wear, power supply fluctuations, and gas pressure changes are difficult to avoid. Traditional control methods cannot comprehensively address these fluctuations, causing process parameters to deviate from set values, resulting in product quality fluctuations and inconsistencies. There is a lack of an intelligent control system capable of sensing parameter fluctuations in real time and automatically performing dynamic compensation. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention proposes a method for controlling atomized powder production using a rotating electrode, comprising the following steps:

[0009] Establish a multivariate machine learning prediction model for the relationship between electrode rotation speed, plasma arc power, and protective gas flow rate with powder particle size and sphericity;

[0010] Based on the particle size and sphericity of the target powder, the optimal process parameters are calculated using the multivariate machine learning prediction model, including: optimal electrode rotation speed, optimal plasma arc power, and optimal protective gas flow rate.

[0011] Apply the optimal process parameters to the powder making control system;

[0012] During the rotary electrode atomization powder production process, the process parameters are monitored and dynamically adjusted in real time to maintain them within the optimal fluctuation range.

[0013] The image acquisition unit acquires images of powder particles in flight, and the particle size and sphericity are analyzed in real time through an AI vision model. Based on the analysis results, the optimal electrode rotation speed, optimal plasma arc power, and optimal protective gas flow rate are adjusted.

[0014] Preferably, the input parameters of the multivariate machine learning prediction model are fused features including the original input and the physical constraint input.

[0015] The initial inputs are: electrode rotation speed, plasma arc power, and protective gas flow rate;

[0016] The physical constraint inputs are: centrifugation term and particle size constraint, sphericity and power and flow rate constraint;

[0017] The original input and the physical constraint input are concatenated to form a fused feature, which is then used as the input parameter for a multivariate machine learning prediction model.

[0018] Preferably, the centrifugation term and particle size constraint are: ;

[0019] Where k1 is an empirical coefficient, ω is the electrode rotation speed, r is the centrifugal radius, and g is the acceleration due to gravity. Indicates the particle size of the powder;

[0020] The sphericity and power and flow constraints are as follows: ;

[0021] Where S is sphericity, P is plasma arc power, and Q is protective gas flow rate. To adjust the parameters.

[0022] Preferably, the steps of analyzing particle size and sphericity in real time using an AI visual model include: (1) removing noise from the image of powder particles using Gaussian filtering or median filtering, separating the powder particles from the background through semantic segmentation, and then extracting individual particles and assigning unique identifiers through contour detection; (2) predicting particle size and sphericity from the image of powder particles using a regression neural network; and (3) updating the particle size histogram and sphericity pass rate in real time.

[0023] Preferably, the steps of adjusting the optimal electrode rotation speed, optimal plasma arc power, and optimal protective gas flow rate based on the analysis results include: increasing the optimal electrode rotation speed when the average particle size is detected to be continuously increasing; increasing the optimal plasma arc power when the average particle size is detected to be continuously smaller than the target particle size range; and adjusting the optimal protective gas flow rate when an increase in satellite powder is detected, leading to a decrease in sphericity.

[0024] Preferably, when the particle size D 50 If the deviation exceeds ±5% or the sphericity pass rate is less than 95%, an alarm is triggered and an adjustment command is sent to the control system via PLC.

[0025] This invention also proposes a rotating electrode atomization powder production control system for implementing the above-mentioned rotating electrode atomization powder production control method. The system is characterized by comprising a sensing system, a data preprocessing module, a model calculation module, an execution control module, and a quality detection module, with each module interacting with data via an industrial Ethernet network.

[0026] The sensing system is a distributed sensor network, which includes a speed sensor installed on the electrode drive shaft, a power sensor integrated into the plasma arc generator, and a gas flow sensor deployed in the protective gas pipeline. All sensor data are aggregated to the data preprocessing module via industrial Ethernet.

[0027] The data preprocessing module performs outlier removal and standardization on the raw data in sequence, and outputs the processed data to the model calculation module.

[0028] The model operation module splits the received data into two paths: one path inputs into the multivariate machine learning model, and the other path stores it in the historical database to support incremental learning; the model operation module can perform forward prediction and backward optimization, and achieve incremental learning by calling the latest data in the historical database, and the operation results are transmitted to the execution control module;

[0029] The execution control module consists of a PLC controller and an actuator. After receiving the instruction, the PLC controller adjusts the electrode speed through the servo driver, adjusts the plasma arc power through the power regulator, and adjusts the protective gas flow through the proportional valve. When the process parameters fluctuate beyond the fluctuation range, the compensation algorithm is automatically started to correct the process parameters.

[0030] The quality detection module includes a high-speed camera and an AI vision processor. The high-speed camera captures images of powder particles, and the AI ​​processor calculates particle size and sphericity through semantic segmentation and feeds the results back to the model calculation module, triggering a multivariate machine learning model to optimize learning parameters.

[0031] Preferably, in the model operation module, during forward prediction, when the single prediction error of particle size exceeds ±4.1μm or the single prediction error of sphericity exceeds ±0.015, historical data with similarity ≥90% in the historical database is automatically called to fine-tune the learning parameters of the multivariate machine learning model.

[0032] Preferably, in the execution control module, the servo driver is driven by a permanent magnet synchronous motor with an electrode speed adjustment accuracy of ±5 rpm. The power regulator has a built-in voltage compensation module to offset the influence of ±10% fluctuation of the grid voltage on the arc power in real time, with a compensation response time of <50 ms. The proportional valve adopts an electro-hydraulic proportional control method with a protective gas flow control accuracy of ±2 SLM.

[0033] Compared with the prior art, the present invention has the following beneficial technical effects:

[0034] 1. This invention establishes a multivariate machine learning prediction model that directly correlates core process parameters such as electrode rotation speed, arc power, and gas flow rate with two key quality indicators: powder particle size and sphericity. This system can not only deduce the optimal process parameters from the target product quality, but also make dynamic feedback adjustments based on real-time monitoring data during the production process. This significantly improves the control accuracy of the final product particle size distribution and sphericity, ensuring the stability and consistency of powder product quality.

[0035] 2. This invention uses centrifugation terms and particle size constraints, as well as sphericity and power flow constraints, as features input into the machine learning model. By integrating the original data with physical constraints, the physical meaning of the model is enhanced, making its decisions not only based on statistical data patterns but also more in line with the physical mechanism of the atomization powdering process. This helps the model converge faster, improves the reliability of predictions in unknown process ranges, and reduces the excessive reliance on training with large amounts of pure data.

[0036] 3. A highly efficient real-time closed-loop feedback control loop was constructed: The system achieves online, non-destructive, real-time detection of powder particles in flight through a high-speed camera and AI vision model. Traditional methods often require interrupting production for offline sieving and microscopic observation, resulting in significant time lag. This invention, through real-time image analysis, can instantly acquire particle size distribution and sphericity information and quickly feed the results back to the control system, promptly suppressing production fluctuations and reducing the scrap rate. Attached Figure Description

[0037] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 This is a flowchart of the rotating electrode atomization powder production control method of the present invention;

[0039] Figure 2 To separate the contour map of each particle through semantic segmentation;

[0040] Figure 3 Image of a metal droplet;

[0041] Figure 4 This is a structural diagram of the model calculation module, execution control module, and quality detection module. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0043] In the accompanying drawings of specific embodiments of the present invention, in order to better and more clearly describe the working principle of each component in the system and show the connection relationship of each part in the device, only the relative positional relationship between each component is clearly distinguished. It does not constitute a limitation on the signal transmission direction, connection sequence, or size and shape of each part within the component or structure.

[0044] Example 1

[0045] like Figure 1 The diagram shows a flowchart of the rotary electrode atomization powder production control method of the present invention, which includes the following steps:

[0046] I. Establish a multivariate machine learning prediction model for the relationship between electrode rotation speed, plasma arc power, and protective gas flow rate with powder particle size and sphericity.

[0047] 1. Data preparation and preprocessing

[0048] To meet the requirement of accurate correlation between process parameters and product performance indicators in the metal powder preparation process, data preparation and preprocessing are carried out first.

[0049] The collected data include: electrode rotation speed ω, plasma arc power P, and protective gas flow rate Q.

[0050] In the centrifugal atomization process for preparing metal powder, the electrode rotation speed, plasma arc power, and protective gas flow rate are the core input variables that determine the powder quality. Among them, the electrode rotation speed directly affects the magnitude of the centrifugal force on the metal droplets. The higher the rotation speed, the more fully the droplets are broken up, and the smaller the particle size is usually.

[0051] The electrode rotation speed ω ranges from 2000 to 15000 rpm, covering different preparation requirements from coarse to fine powder. The plasma arc power determines the melting degree and temperature of the metal raw material. Too low a power will result in incomplete melting of the metal, forming irregular particles, while too high a power will easily cause excessive evaporation of the metal. Therefore, the plasma arc power range is set at 80-200 kW. The protective gas flow rate Q is used to isolate air, prevent metal oxidation, and assist in the cooling of droplets. Insufficient protective gas flow rate will lead to an increase in oxygen content, while too high a flow rate may blow away uncondensed droplets. Therefore, the protective gas flow rate Q ranges from 30 to 100 SLM.

[0052] Regarding the target output variable, particle size D 50 As a key indicator characterizing powder particle size distribution, particle size D directly affects the powder's flowability and formability. 50 The target range is 15-150μm, which can be adapted to different additive manufacturing process requirements; sphericity reflects the regularity of powder particles. High sphericity powder can improve the density of molded parts. The sphericity needs to be controlled between 0.85 and 0.99.

[0053] To ensure the generalization ability and reliability of the prediction model, the dataset must cover at least 500 sets of experimental records and cover a variety of commonly used metal materials such as Ti6Al4V and 316 stainless steel. Among them, Ti6Al4V has excellent high-temperature strength and corrosion resistance, while 316 stainless steel has good biocompatibility and corrosion resistance.

[0054] 2. Data cleaning and standardization

[0055] During the experiment, factors such as equipment operational stability, measuring instrument accuracy, and operating environment may lead to extreme values ​​and noise in the data set. Examples include instantaneous out-of-range fluctuations in electrode rotation speed due to sudden equipment failure, abnormal peaks in plasma arc power caused by voltage instability, and abnormal data in protective gas flow rate due to valve control errors. Directly using this data for modeling would severely deviate from the actual process, resulting in decreased prediction accuracy. Therefore, statistical methods are needed to filter the data, removing extreme values ​​and noise beyond a reasonable range, and normalizing the input and output variables to the [0,1] range. Normalization eliminates the influence of dimensions, ensuring all variables are on the same order of magnitude. This guarantees that the prediction model can uniformly learn the influence of each variable during gradient descent, improving training efficiency and the stability of the prediction model.

[0056] The processed dataset is randomly divided into training, validation, and test sets in a 6:2:2 ratio. The training set (60%) is used for learning and fitting the prediction model parameters and must contain a sufficiently rich combination of process parameters to ensure that the prediction model can fully learn the correlation between input and output. The validation set (20%) is mainly used for adjusting and optimizing the prediction model's learning parameters. By monitoring the performance metrics of the validation set, overfitting of the prediction model can be avoided. The test set (20%) must be kept independent of the training and validation sets and is used to simulate real-world application scenarios to objectively evaluate the generalization ability and prediction accuracy of the prediction model. Random sampling methods must be used during the partitioning process, and the distribution of input and output variables in each dataset must be consistent with the original dataset to avoid distortion of evaluation results due to data distribution deviations.

[0057] 3. Obtain physical constraint relationships by combining univariate modeling.

[0058] To further improve the physical rationality and prediction accuracy of the prediction model, it is necessary to carry out univariate modeling and physical relationship calibration, establish the basic physical correlation between process parameters and output indicators, and provide a constraint basis for subsequent multivariate prediction models.

[0059] By combining univariate modeling to obtain physical constraint relationships, two physical constraint features are constructed: the first constraint feature is the correlation index between centrifugation term and particle size. Based on the empirical coefficients in the univariate model, the electrode rotation speed is converted into a centrifugation term feature directly related to particle size. The second constraint feature is the correlation index between sphericity and power and flow rate. Based on the adjustment parameters in the sphericity model, the plasma arc power and the protective gas flow rate are integrated into a comprehensive feature to characterize the synergistic effect of the two on sphericity.

[0060] (1) Modeling the physical constraints of centrifugation and particle size.

[0061] Particle size

[0062] in, ω is the particle size, r is the electrode rotation speed, g is the centrifugal radius, and k1 is the gravitational acceleration.

[0063] In the process of preparing metal powder by centrifugal atomization, the breakup and particle size formation of metal droplets are mainly driven by centrifugal force. According to physical principles, the magnitude of centrifugal force is directly proportional to the square of the electrode rotation speed and the electrode radius, and inversely proportional to the gravitational acceleration.

[0064] make We fit y=k1x through linear regression and solve for the empirical coefficient k1.

[0065] Preferably, 100 sets of experimental data for Ti6Al4V alloy were selected for fitting. Ti6Al4V exhibits stable physical properties such as melting point and density, resulting in high repeatability and reliability of the experimental data. The least squares method was employed during fitting, minimizing the sum of squared residuals between the actual and predicted particle sizes to determine the optimal empirical coefficient k1. This empirical coefficient quantitatively reflects the influence of the centrifugation term on particle size in the centrifugal atomization process, providing crucial physical constraints for subsequent multivariate model particle size predictions. This avoids predictions that contradict physical laws, such as particle size increasing with increasing rotational speed. Finally, the fitted data... .

[0066] (2) Modeling is performed to address the physical constraints of sphericity S, plasma arc power P, and protective gas flow rate Q:

[0067]

[0068] in, To adjust the parameters, tanh is the hyperbolic tangent function.

[0069] The plasma arc power determines the melting temperature and residence time of the molten metal droplets. If the power is too low, the metal material will not melt completely, and the droplets will not form a regular spherical shape during cooling. If the power is too high, the droplets will overheat, leading to evaporation or splashing, which also affects sphericity. The protective gas flow rate affects sphericity by controlling the cooling rate and isolating the droplets from air. Insufficient flow rate results in slow cooling, making the droplets susceptible to oxidation and forming irregular surfaces. Excessive flow rate increases the impact force of the airflow on the droplets, potentially causing deformation. Adjustable parameters are introduced. To balance the interaction between the two, adjust the parameters. The determination of the optimal value requires optimization through multiple sets of cross-experiment data.

[0070] 3. Construction of multivariate machine learning prediction models

[0071] Based on the preliminary data processing and physical constraint results, a multivariate machine learning prediction model is constructed. By fusing the original input and the physical constraint input, the model's accuracy in predicting particle size D is improved. 50 With the accuracy of sphericity prediction.

[0072] (1) Construction of fusion features

[0073] Raw inputs: electrode rotation speed, plasma arc power, protective gas flow rate.

[0074] Physical constraint input: Based on the results of the univariate model, two physical constraints are input: centrifugation term and particle size constraint, and sphericity and power and flow rate constraint.

[0075] Fusion features: The original input and the physical constraint input are concatenated to form fusion features, which are used as input parameters for the multivariate machine learning prediction model, namely 3 original variables and 2 physical constraints, for a total of 5 dimensions of input.

[0076] By concatenating three original variables with two physical constraints, a fusion feature of five-dimensional input is formed. This fusion feature retains the parameter information of the original input while introducing constraints based on physical laws, which can guide the prediction model to learn key influencing factors more efficiently and avoid the model from getting stuck in overfitting to irrelevant features.

[0077] (2) Structure of the prediction model

[0078] A multilayer perceptron neural network is used as the prediction model. The multilayer perceptron has a strong nonlinear fitting ability and can effectively handle the complex nonlinear relationship between process parameters and output indicators. It is suitable for complex systems with multiple coupled factors, such as metal powder preparation.

[0079] The prediction model is structured as follows: The input layer has 5 nodes, each corresponding to a 5-dimensional fusion feature, ensuring that all key information can be effectively input into the prediction model; the hidden layer has 2 layers, with the first layer containing 128 neurons and the second layer containing 64 neurons, using a decreasing neuron count structure to progressively extract key information from the input features and reduce interference from redundant features. The first layer captures complex interactions between features through a large number of neurons, while the second layer integrates and compresses key features through a smaller number of neurons; the ReLU activation function is chosen. Compared to the traditional Sigmoid or Tanh functions, the ReLU function can effectively solve the gradient vanishing problem, improve the training efficiency of deep networks, and has lower computational complexity, which helps to accelerate model convergence.

[0080] Preferably, to prevent model overfitting, a double regularization mechanism is introduced:

[0081] Specifically, L2 regularization is adopted with a regularization coefficient of 0.0001. By adding a weight squared term to the loss function, the excessive value of the prediction model weights is limited, thus avoiding the prediction model from over-relying on some features. At the same time, an early stopping mechanism is adopted to monitor the loss value of the validation set in real time during the training process of the prediction model. When the loss value of the validation set no longer decreases for several consecutive rounds, the training is stopped immediately to prevent the model from overfitting to the training set after continued training.

[0082] The loss function is preferably the mean squared error, which effectively measures the deviation between the predicted value and the true value. The optimizer is preferably the Adam optimizer, which combines the advantages of momentum gradient descent and adaptive learning rate. It can adaptively adjust the learning rate of each parameter. Compared with traditional stochastic gradient descent, it has a faster convergence speed and is less likely to get stuck in local optima. The learning rate is set to 0.0005. This learning rate has been tested and adjusted in multiple pre-experiments to ensure that the model converges quickly and avoids training oscillations caused by an excessively large learning rate.

[0083] To determine the optimal combination of learning parameters for the prediction model, a parameter search was conducted. The search scope included the number of hidden layers (1, 2, or 3 layers), the number of neurons per layer (32, 64, or 128), the activation function (ReLU and Tanh functions), and the regularization coefficient (0.0001, 0.001, or 0.01). A grid search method was used to traverse all combinations of learning parameters. The mean squared error of the validation set was used as the evaluation criterion to select the combination of learning parameters with the smallest mean squared error, ensuring that the model has both good fitting ability and excellent generalization performance.

[0084] 4. Model Validation and Performance Evaluation

[0085] The historical dataset used in this evaluation contains 500 complete experimental records of metal powder preparation, covering different combinations of process parameters and various metal materials. The range of input variable values ​​strictly follows actual production requirements. During data partitioning, a combination of random sampling and stratified sampling was used, dividing the dataset into training, validation, and test sets in a 6:2:2 ratio. Stratified sampling ensures that the distribution of input and output variables in each dataset is consistent with the original dataset, avoiding distortion of evaluation results due to data distribution bias. The test set, as a completely independent dataset, did not participate in the model's training and parameter optimization process, and can truly reflect the model's generalization ability in real-world application scenarios.

[0086] The performance evaluation uses two core quantitative indicators: mean absolute error and coefficient of determination, which measure the model performance from the two dimensions of error magnitude and model interpretability, respectively.

[0087] For particle size D 50The prediction model's mean absolute error is controlled within ±4.1μm. The smaller the mean absolute error, the smaller the average deviation between the predicted value and the actual particle size, which can meet the precision requirements for particle size control in industrial production.

[0088] Specifically, in the preparation of Ti6Al4V powder in the aerospace field, the particle size deviation needs to be controlled within 5μm, and this prediction model fully meets the requirements. Meanwhile, the coefficient of determination (R²) for particle size prediction is greater than 0.95. The closer the coefficient of determination 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 patterns, and the prediction results have extremely high reliability. For sphericity prediction, the average absolute error of the prediction model is ±0.015. Sphericity, as an indicator of particle regularity, requires a prediction deviation of no more than 0.02 in industrial production, and the error of this prediction model is completely within an acceptable range. The coefficient of determination (R²) for sphericity prediction is greater than 0.89. Although slightly lower than the R² for particle size prediction, considering that sphericity is affected by more subtle factors such as the cooling rate of molten metal droplets and airflow disturbances, this coefficient of determination is sufficient to demonstrate the model's ability to capture the sphericity variation patterns and can provide accurate sphericity predictions for actual production.

[0089] Preferably, to ensure the reproducibility of the evaluation results, a 5-fold cross-validation method is used. The dataset is divided into 5 non-overlapping subsets. Four subsets are selected as the training set and one subset as the test set each time. The experiment is repeated 5 times, and the average performance index is calculated. The results show that the mean absolute error and coefficient of determination of the 5 experiments fluctuate little, proving that the model's performance is stable and reliable, and it can maintain consistent prediction accuracy across different data subsets. Furthermore, this model is compared with traditional linear regression models and univariate prediction models. Traditional models have limitations regarding particle size D. 50 The predicted R² of the previous model was only around 0.8, and the predicted R² for sphericity was less than 0.75. However, the R² of this model was improved by more than 15% respectively, which fully demonstrates the significant advantage of the multivariate machine learning model that integrates physical constraints in terms of prediction accuracy. This model can provide strong support for parameter optimization and quality control of metal powder preparation processes.

[0090] 2. Based on the particle size and sphericity of the target powder, the optimal process parameters are calculated using the multivariate machine learning prediction model, including: optimal electrode rotation speed, optimal plasma arc power, and optimal protective gas flow rate.

[0091] Input the target powder properties; in one specific embodiment, D 50 =50μm, sphericity S≥0.95, the optimal combination of process parameters is optimized through reverse prediction, and the optimal process parameters are output: optimal electrode speed, optimal plasma arc power and optimal protective gas flow rate.

[0092] By combining real-time monitoring data from online particle size analyzers and temperature sensors in production, the learning parameters of the prediction model are dynamically adjusted, and incremental training is performed using new experimental data to maintain consistency between the prediction model and the actual process.

[0093] 3. Apply the optimal process parameters to the powder making process and run the powder making system.

[0094] In a sealed chamber filled with inert gas, the drive device is activated to make the rotating electrode rotate at high speed. After the speed stabilizes, the coaxial plasma arc is ignited to uniformly melt the end face of the electrode rod to form a liquid film.

[0095] This stage requires ensuring the smooth operation of the rotation, arc, cooling, and gas protection systems, and stabilizing the process parameters to the optimal values ​​calculated in step one, thus creating stable conditions for the uniform formation and atomization of droplets.

[0096] To construct a dynamic compensation system, during the atomization process, a sensor network is needed to monitor the fluctuations in the actual rotation speed of the electrodes, the plasma arc power, and the flow rate of the protective gas in the chamber in real time, so as to reduce the mutual interference between process parameters and improve operational stability.

[0097] The monitoring data is compared with the optimal process parameters, which are the previously calculated optimal electrode speed, optimal plasma arc power, and optimal protective gas flow rate. The control system dynamically fine-tunes the electrode speed, plasma arc power, and protective gas flow rate to compensate for possible fluctuations and ensure that the process is always in the best state.

[0098] Fourth, the image acquisition unit acquires images of powder particles in flight, and the particle size and sphericity are analyzed in real time through an AI vision model. Based on the analysis results, the optimal electrode rotation speed, optimal plasma arc power, and optimal protective gas flow rate are adjusted.

[0099] The image acquisition unit uses real-time high-resolution imaging technology combined with AI vision models to analyze the sphericity and particle size of powder online, enabling rapid quality feedback and optimizing the production process.

[0100] In a preferred embodiment, the AI ​​vision model combines image processing with a deep learning AI model, and the analysis method is as follows:

[0101] The image acquisition unit continuously acquires image sequences of the powder flow and uses a filtering algorithm to reduce image noise.

[0102] Semantic segmentation based on deep learning is used to separate powder particles from the background in an image. Semantic segmentation can better handle complex situations such as particle overlap, adhesion, and uneven lighting, accurately separating the contours of each particle, such as... Figure 2 As shown.

[0103] Contour search is performed on the segmented binary image to identify each independent particle region, and a unique identifier is assigned to each particle.

[0104] For each extracted particle region, particle size distribution calculation and sphericity analysis were performed.

[0105] The system runs continuously, analyzing tens of thousands of particles per second. It updates the particle size histogram and sphericity pass rate in real time.

[0106] Set a quality threshold, preferably when the particle size D 50 An alarm will be triggered if the deviation exceeds ±5% or the sphericity pass rate is below 95%, prompting operational intervention. The ultimate goal is to enable the AI ​​system to perform correlation analysis between particle size quality indicators and electrode rotation speed, plasma arc power, and protective gas flow rate.

[0107] Preferably, if the AI ​​vision model detects a continuous increase in average particle size, it can automatically determine and suggest increasing the electrode rotation speed. As the rotation speed increases, the particle size decreases.

[0108] When the AI ​​vision model detects that the average particle size is consistently smaller than the target particle size range, the optimal plasma arc power should be increased. Insufficient plasma arc power leads to insufficient energy input to the molten metal, resulting in low melting volume or low fragmentation energy during atomization, thus causing the overall particle size of the atomized powder to be smaller. In this case, the optimal plasma arc power should be increased to enhance the energy and melting rate of the melt, allowing the atomization process to produce powder particles that better match the target size.

[0109] When the AI ​​vision model detects an increase in satellite dust and a decrease in sphericity, it can determine and suggest adjusting the gas flow rate to reduce droplet collisions and adhesion. For example... Figure 3 As shown.

[0110] These optimization suggestions can be displayed on the interface for manual confirmation and execution, or automatically sent to the actuators via the PLC controller to achieve fully automated real-time process optimization.

[0111] Based on the quality assessment results, it was decided to adjust the parameters for the next batch of production.

[0112] Example 2

[0113] This invention also proposes a rotating electrode atomization powder preparation control system, which achieves precise control of the metal powder preparation process through a sensing system, a data acquisition module, a model calculation module, an execution control module, and a quality detection module.

[0114] The sensing system acquires key process parameters and environmental data in real time through a distributed sensor network.

[0115] A speed sensor is mounted on the electrode drive shaft, capturing the electrode rotation speed at a sampling frequency of 1kHz with an accuracy of ±10rpm. A power sensor is integrated into the plasma arc generator, monitoring the power output in the range of 80-200kW in real time with an error controlled within ±1kW. A gas flow sensor is deployed in the protective gas pipeline, dynamically recording flow changes from 30-100SLM with a response time of less than 1s. In addition, an oxygen content sensor and a temperature sensor monitor the purity of the inert gas in the chamber and the operating temperature, respectively. All sensor data is aggregated to the data preprocessing module via an industrial Ethernet network.

[0116] The data preprocessing module performs multi-step processing on the collected raw data: first, extreme outliers are removed using the 3σ criterion to eliminate noise caused by instantaneous fluctuations in the equipment; then, Z-Score standardization is used to map each parameter to the [0,1] interval to solve the problem of dimensional differences; finally, the sliding window algorithm is used to extract data trend features to provide stable input for the model calculation module.

[0117] The model computation module receives a dataset in two streams: one stream is fed into a multivariate machine learning model for prediction, and the other stream is stored in a historical database to support incremental learning of the model. The model computation module, based on a machine learning model built on the PyTorch framework, performs dual functions: forward prediction calculates particle size and sphericity using 5-dimensional input, with prediction errors controlled within ±4.1 μm for particle size and ±0.015 for sphericity; backward optimization, based on user-defined target performance, uses a reinforcement learning algorithm to output the optimal combination of process parameters: optimal electrode rotation speed, optimal plasma arc power, and optimal protective gas flow rate.

[0118] The model calculation module performs incremental learning by calling the latest data from the historical database in real time, ensuring that the prediction accuracy is maintained even when the equipment ages or the materials are replaced. The calculation results are transmitted to the execution control module via the OPC UA protocol.

[0119] The execution control module consists of a PLC controller and an actuator. After receiving parameter commands from the model, the PLC adjusts the electrode 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 for all adjustment actions is less than 100ms. When the sensing system detects parameter fluctuations exceeding a threshold, the execution control module automatically activates a compensation algorithm to dynamically correct the execution process parameters and maintain process stability.

[0120] In a preferred embodiment, the quality detection module includes an image acquisition unit.

[0121] like Figure 4 As shown, the image acquisition unit includes a high-speed camera and an AI vision processor.

[0122] High-speed camera: Installed on the powder flow path after the atomization chamber, preferably in the settling section or inert gas circulation loop, it has a sufficient frame rate to capture high-speed flying powder particles and sufficient resolution to clearly present the microscopic morphology of individual particles. The camera captures flying powder particles at a frame rate of 30fps, with LED backlighting to ensure image clarity.

[0123] The AI ​​vision processor extracts particle contours through semantic segmentation, calculates particle size and sphericity in real time, and feeds the detection results back to the model computation module. When the detected value deviates from the target value by more than ±5%, the model is triggered to re-optimize parameters. The AI ​​vision processor uses a GPU accelerator card to accelerate the inference process of the deep learning model to meet real-time requirements.

[0124] Preferably, the quality inspection module also includes a high-brightness, uniform backlight or sidelight source: providing stable and uniform illumination to ensure clear particle image outlines and high contrast, reducing shadows and glare, and facilitating subsequent image processing. Typically, an LED surface light source is used for backlighting to form a bright-field image.

[0125] Preferably, the quality inspection module also includes a sealed optical window: an optical window is opened on the atomization chamber and an anti-pollution design is adopted to prevent powder from adhering and obstructing the field of view.

[0126] The above modules interact via industrial Ethernet to form a control system. The sensing system and data acquisition module provide basic data for the system, the preprocessing module ensures data quality, the model calculation module provides intelligent decision-making, the execution control module realizes physical regulation, and the quality inspection module ensures control effectiveness, together achieving high-precision and intelligent management and control of metal powder preparation.

[0127] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for controlling atomized powder production using a rotating electrode, characterized in that, Includes the following steps: Establish a multivariate machine learning prediction model for the relationship between electrode rotation speed, plasma arc power, and protective gas flow rate with powder particle size and sphericity; The input parameters of a multivariate machine learning prediction model are a fusion of features including the original input and the physical constraint input. The initial inputs are: electrode rotation speed, plasma arc power, and protective gas flow rate; The physical constraint inputs are: centrifugation term and particle size constraint, sphericity and power and flow rate constraint; The original input and the physical constraint input are concatenated to form a fused feature, which is then used as the input parameter for a multivariate machine learning prediction model. The centrifugation term and particle size constraint are: ; Where k1 is an empirical coefficient, ω is the electrode rotation speed, r is the centrifugal radius, and g is the acceleration due to gravity. Indicates the particle size of the powder; The sphericity and power and flow constraints are as follows: ; Where S is sphericity, P is plasma arc power, and Q is protective gas flow rate. To adjust the parameters, tanh is the hyperbolic tangent function; Based on the particle size and sphericity of the target powder, the optimal process parameters are calculated using the multivariate machine learning prediction model, including: optimal electrode rotation speed, optimal plasma arc power, and optimal protective gas flow rate. Apply the optimal process parameters to the powder making control system; During the rotary electrode atomization powder production process, the process parameters are monitored and dynamically adjusted in real time to maintain them within the optimal fluctuation range. The image acquisition unit acquires images of powder particles in flight, and the particle size and sphericity are analyzed in real time through an AI vision model. Based on the analysis results, the optimal electrode rotation speed, optimal plasma arc power, and optimal protective gas flow rate are adjusted.

2. The rotating electrode atomization powder production control method according to claim 1, characterized in that, The steps for real-time analysis of particle size and sphericity using an AI visual model include: (1) removing noise from the image of powder particles using Gaussian filtering or median filtering, separating the powder particles from the background through semantic segmentation, and then extracting individual particles and assigning unique identifiers through contour detection; (2) predicting particle size and sphericity from the image of powder particles using a regression neural network; and (3) updating the particle size histogram and sphericity pass rate in real time.

3. The rotating electrode atomization powder production control method according to claim 1, characterized in that, The steps for adjusting the optimal electrode rotation speed, optimal plasma arc power, and optimal protective gas flow rate based on the analysis results include: increasing the optimal electrode rotation speed when the average particle size is continuously increasing; increasing the optimal plasma arc power when the average particle size is continuously smaller than the target particle size range; and adjusting the optimal protective gas flow rate when an increase in satellite powder leads to a decrease in sphericity.

4. The rotating electrode atomization powder production control method according to claim 3, characterized in that, When the particle size D 50 If the deviation exceeds ±5% or the sphericity pass rate is less than 95%, an alarm is triggered and an adjustment command is sent to the control system via PLC.

5. A rotary electrode atomization powder production control system, used to implement the rotary electrode atomization powder production control method as described in any one of claims 1-4, characterized in that, It includes a sensing system, a data preprocessing module, a model calculation module, an execution control module, and a quality inspection module. Each module interacts with the other via an industrial Ethernet network. The sensing system is a distributed sensor network, which includes a speed sensor installed on the electrode drive shaft, a power sensor integrated into the plasma arc generator, and a gas flow sensor deployed in the protective gas pipeline. All sensor data are aggregated to the data preprocessing module via industrial Ethernet. The data preprocessing module performs outlier removal and standardization on the raw data in sequence, and outputs the processed data to the model calculation module. The model computation module splits the received data into two paths: one path inputs into the multivariate machine learning model, and the other path stores it in the historical database to support incremental learning. The model computation module is capable of forward prediction and backward optimization, and incremental learning is achieved by calling the latest data from the historical database. The computation results are then transmitted to the execution control module. The execution control module consists of a PLC controller and an actuator. After receiving the instruction, the PLC controller adjusts the electrode speed through the servo driver, adjusts the plasma arc power through the power regulator, and adjusts the protective gas flow through the proportional valve. When the process parameters fluctuate beyond the fluctuation range, the compensation algorithm is automatically started to correct the process parameters. The quality detection module includes a high-speed camera and an AI vision processor. The high-speed camera captures images of powder particles, and the AI ​​processor calculates particle size and sphericity through semantic segmentation and feeds the results back to the model calculation module, triggering a multivariate machine learning model to optimize learning parameters.

6. The rotary electrode atomization powder production control system according to claim 5, characterized in that, In the model operation module, during forward prediction, when the single prediction error of particle size exceeds ±4.1μm or the single prediction error of sphericity exceeds ±0.015, historical data with similarity ≥90% in the historical database is automatically called to fine-tune the learning parameters of the multivariate machine learning model.

7. The rotary electrode atomization powder production control system according to claim 5, characterized in that, In the execution control module, the servo driver is driven by a permanent magnet synchronous motor with an electrode speed adjustment accuracy of ±5 rpm. The power regulator has a built-in voltage compensation module to offset the impact of ±10% fluctuation in grid voltage on arc power in real time, with a compensation response time of <50 ms. The proportional valve adopts an electro-hydraulic proportional control method, and the protection gas flow control accuracy is ±2 SLM.

Images