Water atomized iron-based powder particle size closed-loop intelligent control method based on multi-modal perception

Abstract

The application discloses a water atomized iron-based powder particle size closed-loop intelligent regulation and control method based on multi-modal perception, relates to the powder metallurgy technical field, and comprises the following steps: S1, collecting an atomization cone image in real time through a visual sensor and extracting a half cone angle, synchronously acquiring melt temperature and atomization water pressure process parameters; S2, determining a visual confidence degree according to the image quality, determining a process parameter modal confidence degree according to the process parameters, and calculating a two-modal fusion weight, so that a current particle size estimation value is obtained by weighted fusion of particle size estimation based on the half cone angle and particle size estimation based on the process parameters; S3, when the half cone angle change rate exceeds a threshold value, calculating a water pressure pre-adjustment increment, when the particle size deviates from a target, calculating a water pressure correction amount, and superimposing a water pressure adjustment instruction output; S4, comparing the measured half cone angle change amount with an expected cone angle response after adjustment, and correcting a control response coefficient.The application realizes multi-modal fusion particle size closed-loop regulation and control, and improves control reliability under complex working conditions.

Description

Technical Field

[0001] This invention relates to the field of powder metallurgy technology, specifically to a closed-loop intelligent control method for the particle size of water-atomized iron-based powder based on multimodal sensing. Background Technology

[0002] Water atomization powder production is currently the mainstream process for large-scale production of iron-based powders. Molten metal flows through the tundish nozzle and is then impacted, broken, cooled, and solidified into powder particles by a high-pressure water jet. The atomizing water pressure and melt temperature determine the powder particle size and particle size distribution. Stable control of particle size is the core objective of water atomization production, directly affecting the pressing and sintering performance of the finished powder.

[0003] Currently, particle size control on production lines generally relies on offline sieving and manual adjustment based on experience. There is an inherent lag between changes in atomization conditions and the completion of sieving at the collection end to obtain particle size data. During this period, the particle size may deviate from the target range, resulting in a large amount of substandard powder. To shorten this lag, some advanced production lines have introduced a parameter-based closed-loop scheme with melt temperature as an indirect controlled variable, adjusting water pressure according to temperature deviations. However, temperature only reflects the energy state of the molten metal flow and has an indirect mapping relationship with particle size. Control accuracy is limited by the coarseness of this mapping. When nozzle wear, water quality changes, or other factors cause operational drift, the originally calibrated temperature-water pressure correspondence gradually becomes mismatched, leading to a continuous decline in control accuracy after long-term operation. Other solutions deploy visual inspection devices in the water atomization system for process monitoring and anomaly alarms. However, visual signals only remain at the observation level and do not enter the closed-loop control circuit, making active particle size control impossible.

[0004] While deep neural network-based visual closed-loop control schemes have emerged in the field of gas atomization, the gas atomization process lacks the significant water mist and steam obstruction, resulting in a stable and controllable airflow field. This differs significantly from the two-phase flow atomization cone and intermittent steam interference characteristics of water atomization, making direct transfer of related technologies impossible. During water atomization, the high-temperature steam generated by the high-pressure water jet intermittently obstructs the observation window, blurring the visual image and incomplete atomization cone contours. Directly using visual signals affected by steam interference will introduce erroneous estimations.

[0005] To address this, we propose a closed-loop intelligent control method for the particle size of water-atomized iron-based powder based on multimodal sensing. Summary of the Invention

[0006] To address the problems raised in the background art, this invention provides a closed-loop intelligent control method for the particle size of water-atomized iron-based powder based on multimodal sensing, specifically including the following steps:

[0007] S1. Real-time acquisition of atomization cone images via a visual sensor, obtaining current melt temperature and atomization water pressure process parameters, and extracting the atomization cone half-cone angle from the atomization cone image. The visual sensor employs a short-wave infrared thermal imager, utilizing the difference in radiance between high-temperature metal droplets and the background of steam and water mist in the short-wave infrared band to extract the outline of the metal droplets; S1 includes:

[0008] Background subtraction and filtering are performed on the acquired fog cone image, and the fog cone contour boundary is extracted based on the processed image;

[0009] Taking the point where the water jet collides with the molten metal as the vertex, draw tangents along the contours of both sides of the cone, and calculate half of the angle between the two tangents to obtain the half cone angle;

[0010] Calculate the diameter of the cone section at a specified height from the point of collision between the water jet and the molten metal flow to obtain the radial width data of the spray.

[0011] Calculate the coefficient of variation of pixel grayscale distribution within a cross section at a specified height from the collision point of the water jet and the molten metal flow to obtain the density uniformity data of the cone cross section;

[0012] The semi-cone angle, spray radial width data, and density uniformity data are used as feature inputs for the visual modality;

[0013] S2. Determine the visual confidence level based on the quality of the atomized cone image, and determine the process parameter modal confidence level based on the melt temperature and atomized water pressure process parameters. Determine the fusion weight of the visual modal and the process parameter modal based on the visual confidence level and the process parameter modal confidence level. Weight the particle size estimation result based on the half cone angle and the particle size estimation result based on the melt temperature and atomized water pressure to obtain the current particle size estimate. When the visual confidence level is lower than the preset drop trigger threshold, the visual modal fusion weight is reduced to the preset lower limit, the particle size estimation degenerates into a pure process parameter driven mode, and the water pressure pre-adjustment increment output of the fast prediction loop is simultaneously locked.

[0014] When the visual confidence level rises above the preset recovery threshold, the fusion weight gradually recovers at the preset rate, and the water pressure pre-adjustment incremental output of the rapid prediction loop is simultaneously unlocked; the descent trigger threshold is lower than the rise recovery threshold, and the two constitute a hysteresis band.

[0015] S3. When the rate of change of the half-cone angle exceeds the preset rate threshold, calculate the water pressure pre-adjustment increment according to the preset estimated gain and output it. The pre-adjustment increment is locked until the rate of change of the half-cone angle falls back below the rate threshold. When the particle size estimate deviates from the target particle size, calculate the water pressure correction amount according to the preset feedback coefficient. Superimpose the water pressure pre-adjustment increment and the water pressure correction amount to generate a water pressure adjustment command and output it for execution. Calculate the coefficient of variation of the pixel grayscale distribution in the cross section at a specified height from the collision point of the water jet and the molten metal flow. Based on the coefficient of variation, obtain the density uniformity data of the cross section inside the cone. When the density uniformity data of the cross section inside the cone is lower than the preset abnormal threshold, it is determined to be an abnormal atomization condition. Lock the water pressure pre-adjustment increment output of the fast prediction loop, suspend the control response coefficient correction, retain the weak feedback to maintain basic control and trigger an alarm.

[0016] The locking will be automatically released once the uniformity data of the density of the cone's internal cross section is restored.

[0017] S4. After the water pressure adjustment is executed, the measured change in half cone angle is compared with the expected cone angle response under the corresponding pre-stored calibrated water pressure adjustment. When the deviation exceeds the preset correction threshold and the visual confidence is not lower than the preset reliability threshold, the control response coefficient is corrected. The actual particle size data obtained from offline screening is periodically injected into the system to correct the reference offset of the fusion model.

[0018] Preferably, the melt temperature and atomization water pressure process parameters are interpolated and resampled according to the time of the visual frame, so that each visual image corresponds to a set of time-synchronized process parameter values.

[0019] Preferably, the visual confidence level is calculated by combining the contrast of the current frame image, the edge gradient intensity, and the integrity of the fog cone contour;

[0020] The modal confidence level of the process parameters is calculated from the coefficient of variation of melt temperature and atomizing water pressure within the recent sliding window.

[0021] Preferably, upper and lower limits and change rate restrictions are set for the water pressure regulation command; when the difference between the water pressure and the upper or lower limit is less than the preset pressure margin, the estimated gain is reduced accordingly according to the ratio of the difference to the pressure margin.

[0022] Preferably, when the measured change in the half-cone angle is less than the expected change, the absolute value of the control response coefficient is decreased; when the measured change in the half-cone angle is greater than the expected change, the absolute value of the control response coefficient is increased.

[0023] After calibration, the feedback coefficients are updated synchronously to ensure that the feedback coefficients and control response coefficients are in a suitable relationship.

[0024] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0025] The proposed closed-loop intelligent control method for the particle size of water-atomized iron-based powder introduces a visual confidence assessment mechanism. This mechanism dynamically adjusts the weight of each mode in particle size estimation based on the clarity of the atomized cone image. When steam interference is strong or visual quality deteriorates, the dependence on the visual mode is automatically reduced, improving the reliability of particle size estimation results under complex operating conditions. A dual-loop control structure combining rate-triggered feedforward and particle size deviation feedback is employed. When the half-cone angle changes significantly, the feedforward channel responds quickly to water pressure adjustment, while the feedback channel eliminates residual deviations in the steady-state phase. After water pressure adjustment, the control response coefficient is corrected by comparing the actual observed cone angle change with the calibrated expected value, enabling the control model to adapt to long-term changes such as nozzle wear and operating condition drift, maintaining the continuity of control effectiveness. By setting a confidence degradation and prediction loop lockout mechanism, the feedforward output is paused, the derating weak feedback channel is retained, and an alarm is triggered when visual information is severely unavailable, avoiding erroneous control commands caused by low-quality visual data. Attached Figure Description

[0026] Figure 1 This invention relates to a closed-loop intelligent control method for the particle size of water-atomized iron-based powder. Detailed Implementation

[0027] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0028] Before execution, the system startup phase loads the pre-stored initial values ​​of various control parameters, including the initial value of the control response coefficient, the estimated gain, the feedback coefficient, the rate threshold, the pressure margin, and the correction threshold. These initial values ​​are obtained in advance through offline calibration experiments and stored in non-volatile memory.

[0029] The method described in this embodiment is based on dual closed-loop control, including a fast prediction loop and a particle size feedback loop. Specifically, the fast prediction loop is a feedforward control loop based on the rate of change of the half-cone angle, and the particle size feedback loop is a feedback control loop based on the particle size estimation error.

[0030] The rapid prediction loop takes the rate of change of the atomizing cone's semi-cone angle as input and suppresses particle size fluctuations caused by sudden changes in operating conditions through water pressure pre-adjustment. The particle size feedback loop takes the deviation between the estimated particle size and the target particle size as input and eliminates steady-state errors through water pressure correction. The outputs of the two loops are superimposed to generate a water pressure regulation command.

[0031] Reference Figure 1 As shown, the closed-loop intelligent control method for particle size of water-atomized iron-based powder based on multimodal sensing includes the following steps:

[0032] S1. Real-time acquisition of atomizing cone images, as well as current melt temperature and atomizing water pressure process parameters, through a visual sensor, and extraction of the atomizing cone half-cone angle from the atomizing cone images.

[0033] Background subtraction and filtering are performed on the acquired atomized cone images. Background subtraction uses inter-frame differencing, suitable for scenarios with dynamic background changes such as splashing water droplets and steam disturbances during water atomization, and can adaptively track the slow evolution of the background. When the proportion of residual pixels judged as foreground changes within the entire effective area after inter-frame differencing exceeds the preset steam disturbance criterion, the system automatically switches to Gaussian mixture modeling to further suppress dynamic background interference. In a specific example, the steam disturbance criterion is set to 15% to 25% of the total effective area of ​​the entire image, corresponding to a condition where steam is diffused but a static background is still discernible; the Gaussian mixture model has 2 to 3 components, corresponding to the static background, slowly changing steam, and splashing water droplets, respectively.

[0034] The filtering process employs an adaptive median filtering method, where the filter kernel size must balance noise suppression capability with edge preservation accuracy. A kernel size that is too small cannot effectively filter out impulse noise from splashing water droplets, while a kernel size that is too large will blur the edges of the fog cone contour. Based on the ambient noise level, the filter kernel size is adaptively selected within the range of 3×3 to 7×7 according to the noise intensity. The fog cone contour boundary is then extracted from the processed image.

[0035] Because high-temperature molten metal exhibits significant thermal radiation in the 1.0 to 1.7 μm wavelength range in short-wave infrared images, while room-temperature water jets and the background are close to ambient temperature and exhibit extremely weak thermal radiation, the molten metal appears as a high-grayscale bright area in short-wave infrared images, forming a significant contrast with the dark background. This grayscale difference is utilized to extract the outline of the molten metal column and the edge contours of the two sides of the atomized cone formed after collision through threshold segmentation. Furthermore, the central axis of the molten metal column is obtained by fitting the contours of the two sides of the molten metal column. Then, the tangents of the two sides of the atomized cone are extended upwards, and the intersection of the tangents with the central axis is used as the initial coarse localization result of the collision intersection point.

[0036] Specifically, the pixel position with the largest gradient magnitude is searched within a preset pixel neighborhood centered on the intersection point. This position is then precisely located as the collision intersection point of the molten metal flow and the high-pressure water jet using bilinear sub-pixel interpolation. This location serves as the geometric reference for subsequent calculations of the semi-cone angle, spray radial width, and density uniformity. In one optional scheme, the preset pixel neighborhood radius is set to 3 to 5 pixels. Too small a radius may miss true edge points, while too large a radius may introduce false edge points from irrelevant areas.

[0037] Using the point of intersection between the water jet and the molten metal flow as the vertex, tangents are drawn along the contours of both sides of the cone. The effective contour segment selected for tangent fitting must avoid the nonlinear region where the flow converges near the cone apex; the upper-middle segment below the collision intersection point, where the contour has become nearly straight, is selected for fitting. The RANSAC robust fitting method is used to construct the tangent equations for both contours. In a specific example, with a resolution of 2048×2048 and a field of view of approximately 200mm, the effective contour segment is approximately 15 to 30 pixels from the collision intersection point, corresponding to an actual physical distance of approximately 15 to 30mm. The convergence criterion is an inlier residual of less than 2 pixels, and the number of iterations does not exceed 100. Half the angle between the two tangents is calculated to obtain the half-cone angle.

[0038] Calculate the diameter of the cone cross-section at a specified height from the collision point of the water jet and the molten metal flow to obtain the radial width of the spray. The selection of the specified height must satisfy the following conditions: a section in the middle of the cone with a relatively gentle change in radial width and a relatively uniform density distribution within the cross-section should be chosen to avoid areas near the cone apex where the liquid flow is not yet fully developed and areas far from the cone apex where the spray is excessively diffused. In one optional configuration, taking a 2048×2048 resolution and a field of view of approximately 200mm as an example, the specified height corresponds to an actual physical distance of approximately 100 to 150mm from the collision point. For atomizing cones of different specifications, the height is adaptively adjusted according to a fixed proportion of the total cone height based on the geometric similarity of the cones.

[0039] The coefficient of variation of pixel grayscale distribution within a cross-section at a specified height from the collision point of the water jet and the molten metal flow is calculated to obtain the density uniformity data of the cone cross-section. The coefficient of variation... Defined as the ratio of the standard deviation σ of pixel grayscale within a cross-section to the mean grayscale value μ, i.e. Since a smaller coefficient of variation indicates a more concentrated grayscale distribution and a more uniform droplet density distribution within the spray cross-section, the cross-sectional grayscale coefficient of variation is negatively correlated with the density uniformity within the cone. The semi-cone angle, spray radial width, and density uniformity together serve as feature inputs for the visual modality.

[0040] The melt temperature and atomizing water pressure process parameters are interpolated and resampled according to the time of the visual frame, using a cubic spline interpolation method. When the sampling frequency of the process parameters is much higher than the visual frame rate, the data changes between adjacent sampling points are small, and linear interpolation can meet the accuracy requirements. However, when the sampling rate is close to the visual frame rate, the cubic spline can give full play to its smoothness advantage and avoid the piecewise angle effect introduced by linear interpolation, so that each frame of visual image corresponds to a set of time-synchronized process parameter values.

[0041] The visual sensor employs a short-wave infrared thermal imager, utilizing the difference in radiance between high-temperature molten metal droplets and the background of steam and water mist in the short-wave infrared band to extract the outline of the molten metal droplets. The physical basis for this is that in the 1.0 to 1.7 μm band, the absorption coefficient of liquid water reaches tens of cm⁻¹, while the absorption coefficient of gaseous water (steam) is 2 to 3 orders of magnitude lower. Therefore, the thermal radiation in this band mainly originates from high-temperature molten metal droplets and liquid water droplets at temperatures of approximately 1500 to 1700°C, with the radiation contribution of steam itself being negligible. Simultaneously, the thermal radiation brightness of the molten metal droplets is approximately 10 to 100 times that of steam at the same temperature, thus forming a high-contrast droplet outline and a low-contrast steam background in the image. The image acquired by the camera focuses on the molten metal droplet outline, rather than the outer boundary of the water jet. In a specific example, the thermal imager operates in the 1.0 to 1.7 μm band, with a spatial resolution of 2048×2048 pixels and a frame rate of no less than 10 fps to meet the second-level timescale requirements of changing operating conditions.

[0042] The camera is mounted in a top-view or oblique-down view relative to the atomization area, with the angle between the optical axis and the axis of the atomization cone, for example, ranging from 30° to 60°, and the field of view fully covering the section from the collision point to the fully developed atomization cone. The optical window is equipped with a sealed protection device, and a reference calibration spot with a known gray level is set on the outside of the window as a quantitative benchmark for the degree of contamination. The average gray level of the reference calibration spot area is continuously monitored during system operation.

[0043] When the grayscale offset exceeds a preset contamination threshold, the contamination level is deemed excessive, triggering automatic purging and cleaning. The contamination threshold is set based on the fact that when the grayscale offset caused by deposits on the window exceeds this threshold, the image contrast and edge sharpness have significantly decreased, affecting the extraction accuracy of cone angles and contour features. In one optional scheme, the contamination threshold is set to 15 to 30 grayscale levels (based on an 8-bit grayscale image), and the purging air source pressure is slightly higher than the pressure inside the atomization tower, for example, 0.02 to 0.05 MPa higher, to ensure effective removal of water mist and splashes adhering to the window surface. During the transition phase where the contamination level exceeds the threshold but purging has not yet been triggered, the system automatically lowers the upper limit of visual confidence to reduce the trust weight of the visual data.

[0044] Specifically, the upper limit of visual confidence The grayscale value is linearly reduced based on the ratio of the grayscale offset to the contamination threshold. When the grayscale offset approaches the contamination threshold, the upper confidence limit approaches zero, triggering automatic purging. The melt temperature sensor uses an armored K-type thermocouple, installed on the side wall below the molten metal surface in the tundish, with a sampling frequency of 10 to 50 Hz as an example. The atomizing water pressure sensor is installed on the pipeline section between the high-pressure water pump outlet and the nozzle, with a sampling frequency no less than twice the visual frame rate, corresponding to the visual frame time after interpolation and resampling.

[0045] S2. Determine the visual confidence level based on the quality of the atomized cone image, determine the process parameter modal confidence level based on the melt temperature and atomized water pressure process parameters, determine the fusion weight of the visual modal and the process parameter modal according to the visual confidence level and the process parameter modal confidence level, and weight and fuse the particle size estimation results based on the half cone angle with the particle size estimation results based on the melt temperature and atomized water pressure to obtain the current particle size estimate.

[0046] Visual confidence is calculated by combining the contrast of the current frame image, edge gradient intensity, and fog cone contour integrity. Specifically, the contrast is quantized using the contrast features of the image's gray-level co-occurrence matrix, the edge gradient intensity is calculated using the Sobel operator to calculate the mean of the gradient magnitudes across the entire frame, and the contour integrity is calculated as the ratio of the actual extracted effective contour arc length to the theoretical complete cone contour arc length corresponding to the same cone angle with the collision intersection as the vertex.

[0047] The determination of the effective contour arc length is based on the principle of geometric consistency, specifically:

[0048] For each extracted contour point, the deviation angle between the gradient direction at that point and the theoretical generatrix direction with the collision intersection as the vertex is calculated. When the deviation angle is less than a preset angle tolerance, the point is marked as a valid contour point. The angle tolerance is determined by the root sum of the squares of three error components: image discretization error, physical deviation of liquid film surface fluctuation, and contour offset caused by slight vapor disturbance. For example, at a resolution of 2048×2048, the image discretization error is approximately 0.3°, the physical deviation of liquid film surface fluctuation is approximately 0.8°, and the contour offset caused by slight vapor disturbance is approximately 0.5°. The typical value of the synthesized angle tolerance is 1.0°~1.5°.

[0049] The sum of the arc lengths formed by all consecutive valid contour points is the effective contour arc length.

[0050] This determination is based on the geometric shape features of the contour rather than the grayscale quality of the image, and is therefore independent of the two sub-indicators of contrast and gradient strength.

[0051] The three sub-indicators are normalized to the [0,1] interval and then summed according to preset weights to obtain the visual confidence score. :

[0052]

[0053] in, The normalized value of image contrast is calculated using the contrast features of the gray-level co-occurrence matrix and then normalized to the [0,1] interval. The normalized value of the edge gradient intensity is calculated by using the Sobel operator to calculate the mean of the gradient magnitude of the entire frame and then normalized to [0,1]. The normalized value of the fog cone contour integrity is equal to the actual extracted effective contour arc length divided by the theoretical complete cone contour arc length corresponding to the same cone angle with the collision intersection as the vertex, normalized to [0,1]. The objective judgment standard of the effective contour arc length is defined based on geometric consistency rather than image quality itself. , , These are the weights of the normalized values ​​for image contrast, edge gradient intensity, and fog cone contour integrity, respectively. .

[0054] Each weight was determined through offline calibration experiments. Specifically, under typical operating conditions of light, medium, and heavy steam disturbance, sufficient image samples were collected. Using the manually calibrated true value of the half-cone angle as a benchmark, each sub-index was individually perturbed, and the corresponding half-cone angle extraction error was statistically analyzed. Normalization was performed using the reciprocal of the proportion of each sub-index error to the total error to obtain the corresponding weight. Contrast and contour integrity have a more direct impact on the reliability of feature extraction, and their weights are slightly higher than those of edge gradient strength. In one optional scheme... Take a value between 0.35 and 0.45. Take a value between 0.15 and 0.30. Take a value between 0.35 and 0.45.

[0055] The modal confidence level of the process parameters is calculated from the coefficients of variation of melt temperature and atomizing water pressure within the recent sliding window. The more stable the process parameters, the higher the confidence level; that is, the confidence level and the coefficient of variation are negatively correlated. Specifically, the coefficient of variation of melt temperature within the sliding window is calculated separately. Coefficient of variation of water pressure Each coefficient of variation is mapped to a confidence value in the interval (0,1) using the single-parameter confidence formula. :

[0056]

[0057] in, The coefficient of variation for temperature or water pressure parameters. These are temperature or water pressure parameters. The mapping function is a monotonically decreasing function; the confidence level is 1 when the coefficient of variation is zero, and it monotonically approaches zero as the coefficient of variation increases. Furthermore, it is less sensitive to small random fluctuations than the linear mapping form, thus fitting the distribution characteristics of parameter fluctuations under industrial conditions. The modal confidence level of the process parameters is then obtained by equal-weighted averaging. :

[0058]

[0059] In the formula, The confidence level for the melt temperature. The confidence level of the water pressure;

[0060] The selection of the sliding window length needs to balance the statistical stability of the coefficient of variation estimation with the response speed to operating condition changes. A window length that is too short will make the coefficient of variation estimation results susceptible to random fluctuations and lack statistical stability; a window length that is too long will result in a delayed response to operating condition changes and an inability to reflect parameter fluctuations in a timely manner. In a specific example, the sliding window length is selected to be 20 to 50 sampling periods.

[0061] During water atomization, the mapping relationship between the semi-cone angle of the atomizing cone and the powder particle size has a clear physical mechanism. An increase in the semi-cone angle means that the impact kinetic energy of the water jet on the molten metal flow is enhanced, the flow is broken up more completely, and the particle size is reduced.

[0062] However, once the semi-cone angle increases to a certain extent, the liquid flow is already sufficiently broken up, and the marginal effect of further increasing the semi-cone angle on particle size gradually weakens, exhibiting a saturation effect. This monotonically decreasing + asymptotically saturated physical characteristic naturally matches the function shape of a second-order polynomial; the first term captures the linear decreasing trend in the kinetic energy-dominated region, while the quadratic term characterizes the asymptotic characteristics of the breaking-up saturation effect. Compared to black-box models such as neural networks, the second-order polynomial contains only 3 parameters to be identified. Under the sample size conditions of industrial online identification, more than 50 sets of effective data can be accumulated within minutes after each operating condition switch to obtain statistically reliable fitting results, while neural networks usually require hundreds to thousands of samples to avoid overfitting.

[0063] Meanwhile, the computational cost of a single frame of the polynomial model is in the microsecond range, which meets the requirements of real-time control for millisecond-level response.

[0064] The effects of melt temperature and atomizing water pressure on particle size are different. Temperature affects droplet breakage by changing the viscosity and surface tension of the molten metal. Within the normal production temperature range of 1500°C to 1700°C, the effect of viscosity change on particle size is approximately linear.

[0065] Water pressure affects the degree of crushing by altering the impact kinetic energy, and this effect is approximately linear within the operating range. Therefore, multiple linear regression is employed to capture the independent linear contributions of the two process parameters to particle size. In terms of real-time performance, the computational cost per frame of linear regression is significantly lower than that of kernel methods such as support vector regression, making it suitable for the response time constraints of closed-loop control. Regarding interpretability, the regression coefficients b1 and b2 have clear physical meanings (temperature sensitivity and pressure sensitivity), facilitating understanding and verification by engineers.

[0066] The median particle size D50 of a powder is defined as the equivalent spherical particle size corresponding to a cumulative particle size distribution percentage of 50%. The particle size estimation model based on the half-cone angle adopts a second-order polynomial form:

[0067]

[0068] in, The median particle size of the powder is estimated based on the visual characteristics of the half-cone angle. The semi-cone angle of the atomizing cone. , , These are the model coefficients; in one alternative approach, the exemplary range of values ​​for the model coefficients is as follows: The range is 120 to 180 μm, reflecting the basic particle size level and corresponding to the typical particle size under medium atomization intensity. The range is -3.0 to -1.5 μm / °, which reflects the linear sensitivity of the half-cone angle to the particle size. A negative value indicates that the particle size decreases as the half-cone angle increases. The range is from -0.05 to 0 μm / °², reflecting the correction magnitude for the saturation effect.

[0069] In the corresponding training samples, the coverage range of the semi-cone angle α is 15° to 45°, the coverage range of D50 obtained by offline screening is 50 to 200 μm, and the sample collection time span covers at least 3 different water pressure conditions to ensure the generalization of the model.

[0070] During water atomization, a larger semi-cone angle of the atomizing cone indicates a stronger impact kinetic energy effect of the water jet on the molten metal flow, more complete flow breakup, and a smaller final powder particle size. The semi-cone angle is related to the particle size... They exhibit a stable, monotonically negative correlation, and this physical mechanism ensures the monotonicity of the polynomial model in this interval;

[0071] The model coefficients were obtained by collecting no fewer than 50 sets of half-cone angles and corresponding offline sieved particle size samples covering the actual water pressure regulation range under stable operating conditions, and minimizing the loss function using the least squares method. The fitting yielded, where The median particle size is the true value obtained from the offline screening of the k-th group. This represents the measured value of the half-cone angle at the corresponding time. The total number of samples and .

[0072] The particle size estimation model based on melt temperature and atomizing water pressure adopts a multiple linear regression approach:

[0073]

[0074] in, The median particle size of the powder is estimated based on melt temperature and atomizing water pressure; T is the melt temperature, and P is the atomizing water pressure. , , For regression coefficients. In one alternative approach, the exemplary range of values ​​for the regression coefficients is as follows: Use 250 to 400 μm as the reference particle size intercept; The range is -0.3 to -0.1 μm / °C, which reflects the sensitivity of temperature to particle size. A negative value indicates that as the temperature increases, the viscosity of the molten metal decreases and the droplets are more easily broken, resulting in a smaller particle size. The range is -5.0 to -2.0 μm / MPa, reflecting the sensitivity of water pressure to particle size. A negative value indicates that the increased impact kinetic energy leads to a decrease in particle size when the water pressure increases. In the corresponding training samples, the melt temperature T covers an exemplary range of 1500 to 1700°C, and the atomizing water pressure P covers an exemplary range of 8 to 20 MPa.

[0075] The regression coefficients are obtained by minimizing the loss function using historical operating data. Obtain, among which The median particle size is the true value obtained from the offline screening of the k-th group. and These represent the melt temperature and atomizing water pressure at the corresponding moments. The model's applicability is limited by the temperature and atomized water pressure ranges covered by the training data, representing the total number of historical data samples.

[0076] If the number of calibration samples is less than 50, the system will run with the preset default model coefficients and continue to accumulate samples. Once the number of samples reaches 50, the system will automatically trigger online calibration to update the model coefficients.

[0077] The fusion weights are obtained by normalizing the two-modal confidence scores by proportion:

[0078]

[0079]

[0080] In the formula, The weights of visual modalities in particle size fusion; The weights of process parameter modes in particle size fusion; satisfying Final particle size estimate for:

[0081]

[0082] Baseline offset parameters of the fusion model This is used to correct systematic biases caused by systematic errors and model approximations; the corrected final particle size estimate is... The initial value of the baseline offset parameter is zero, and it is updated by the subsequent offline screening true value calibration step.

[0083] Furthermore, when the visual confidence level falls below a preset drop trigger threshold... At this time, the visual modality fusion weights are lowered to a preset lower limit to quickly shield the interference of low-quality visual data on particle size estimation. Particle size estimation degenerates into a pure process parameter driven mode, and the water pressure pre-adjustment increment output of the fast prediction loop is simultaneously locked. When the visual confidence recovers to the preset rise recovery threshold, When the above occurs, the fusion weight gradually recovers to the normal value at a preset rate. The typical value of the recovery rate is an increase of 0.05 to 0.15 weight percentage per second, which can be determined through offline debugging according to the stability requirements of the working conditions. The smaller the rate, the smoother the recovery and the higher the control stability; the larger the rate, the faster the recovery response.

[0084] The rapid prediction loop's water pressure pre-adjustment incremental output synchronously releases the interlock. Among other things... The two constitute a hysteresis band, which avoids frequent weight jumps that cause control oscillations when the confidence level fluctuates near the threshold.

[0085] The basis for the asymmetric design, which jumps directly when the weight decreases and gradually changes when it recovers, is as follows:

[0086] From an information theory perspective, a sudden drop in visual confidence indicates a sudden change in image quality, such as a vapor cloud momentarily obscuring the image. At this time, the information entropy of the visual data increases sharply, and the reliability drops to an unusable level within a single frame. Therefore, an immediate jump is required to reduce the risk of incorrect estimation.

[0087] Visual confidence recovery is a gradual process; the vapor dissipates slowly, and image quality improves progressively. The recovery of information content is continuous. If the weights were to undergo a step recovery simultaneously, it would be equivalent to assigning abruptly high weights to data that has not yet been fully validated, potentially introducing new estimation biases. Therefore, abrupt descent provides rapid protection, while gradual recovery ensures a smooth transition.

[0088] The preset reliability threshold is the lowest visual modal confidence level at which visual data can support the quantization correction of the control response coefficient. It is set based on the fact that when the visual modal confidence level is below this threshold, the quantitative error of cone angle extraction is too large to support accurate coefficient correction. The preset reliability threshold is not lower than the confidence rise recovery threshold, exemplarily set to 0.5~0.6. The control response coefficient correction operation in S4 is only allowed when the visual quality recovers to a reliable level. In an optional scheme, a fall trigger threshold... Take a value between 0.3 and 0.4 to increase the recovery threshold. The weights are set to 0.4 to 0.55, and the hysteresis band width is set to 0.1 to 0.15, balancing anti-fluctuation performance and recovery sensitivity. The lower limit of the visual fusion weights is set to 0.1. The basis for setting the lower limit of the weights is to retain a minimum visual constraint even when the visual quality is poor, so as to avoid the system completely degenerating into a single-modal estimation and losing the advantages of multimodal redundancy.

[0089] S3. When the rate of change of the half-cone angle exceeds the preset rate threshold, calculate and output the water pressure pre-adjustment increment based on the preset estimated gain. This pre-adjustment increment is locked until the rate of change of the half-cone angle falls back below the rate threshold. When the estimated particle size deviates from the target particle size, calculate the water pressure correction amount based on the preset feedback coefficient. Add the water pressure pre-adjustment increment and the water pressure correction amount to generate a water pressure regulation command and output it for execution.

[0090] The rate of change of the half-cone angle is calculated using the slope obtained by linear fitting of the half-cone angle time series within a sliding window. The selection of the window length needs to balance the accuracy of the slope estimation and the sensitivity to sudden changes in operating conditions: if the window is too short, the slope estimation is easily affected by inter-frame noise and the estimation is unstable; if the window is too long, the response to sudden changes in operating conditions is lagging. In one optional scheme, the window length is selected as 5 to 10 frames.

[0091] When the rate of change of the half-cone angle exceeds a preset rate threshold, the water pressure pre-adjustment increment is calculated based on the estimated gain and output. The estimated gain is the water pressure adjustment required to generate a unit steady-state change in the half-cone angle, with dimensions in MPa / °. It is obtained through pre-calibration: under steady-state conditions, a step change of known amplitude is applied to the water pressure; after the response stabilizes, the steady-state change in the half-cone angle is recorded; the estimated gain is calibrated using the ratio of the water pressure adjustment to the steady-state change in the half-cone angle. Water pressure pre-adjustment increment. The calculation formula is:

[0092]

[0093] in To estimate the gain, The rate of change of the half-cone angle is defined as follows: the rate of change corresponding to the half-cone angle fluctuation under normal operating conditions is much lower than this threshold, while the rapid change of the cone angle caused by sudden changes in nozzle operating conditions significantly exceeds this threshold, and there is a clear distinction between the two. In one optional scheme, the rate threshold is set to 0.5° / s to 2° / s.

[0094] The pre-adjustment increment locking refers to the following: when the rate of change of the half-cone angle exceeds a rate threshold, the system continuously calculates the water pressure pre-adjustment increment at the current rate. As long as the increment does not exceed the preset upper limit of the pre-adjustment increment amplitude, it updates in real time with the rate. When the increment reaches the upper limit, it locks at the upper limit and no longer increases until the release condition is met. If the trend of the half-cone angle change reverses during the locking period, the lock is immediately released and the pre-adjustment increment is cleared to zero. The upper limit of the pre-adjustment increment amplitude is used to constrain the maximum allowable water pressure pre-adjustment amplitude under a single operating condition disturbance. Exceeding this amplitude may cause the water pressure adjustment to exceed the safe operating range of the actuator or cause system overshoot oscillation. In one optional scheme, this upper limit is set to a single water pressure adjustment amount not exceeding 5% to 10% of the current operating pressure.

[0095] When the rate drops below the rate threshold, the lock is released, and the pre-adjustment increment is gradually withdrawn using a ramp function instead of being cleared by a step. The ramp withdrawal duration needs to balance control stability and reset speed to avoid system oscillations caused by sudden changes in control input. In one optional scheme, the ramp withdrawal duration is 1 to 3 seconds.

[0096] When the estimated particle size deviates from the target particle size, the water pressure correction is calculated according to a preset feedback coefficient. Feedback coefficient The physical meaning of is the water pressure adjustment required to generate a unit particle size correction, with dimensions in MPa / μm, and it maintains a preset fit with the control response coefficient. Water pressure correction The calculation formula is:

[0097]

[0098] in, The difference between the corrected particle size estimate and the target particle size is... ; The preset target powder median particle size; This is the final particle size estimate after incorporating the reference offset. By introducing the reference offset parameter ΔD0, the feedback loop can utilize the cumulative error information from offline sieving correction, avoiding the inability to eliminate systematic biases caused by relying solely on real-time visual estimation. The target particle size is preset according to the product specifications of the powder to be produced.

[0099] The water pressure pre-adjustment increment and the water pressure correction amount are added together to generate a water pressure adjustment command and output it for execution.

[0100]

[0101] In the formula, This setting determines the final water pressure command value output to the actuator; it also sets upper and lower limits and a rate of change limit for the water pressure adjustment command. The upper limit must ensure safe operation of the equipment and allow for a pressure margin; for example, it should not exceed 90% of the rated operating pressure of the atomization system. The lower limit must ensure sufficient water pressure at the nozzle to avoid poor atomization; for example, it should be 105% to 110% of the minimum pressure required for normal nozzle atomization. The rate of change limit is used to prevent system oscillation caused by sudden changes in water pressure; for example, a single adjustment should not exceed 5% of the current water pressure value.

[0102] When the difference between the water pressure and the upper or lower limit is less than the preset pressure margin, the estimated gain is reduced accordingly based on the ratio of the difference to the pressure margin. The formula is:

[0103]

[0104] in This is the difference between the current water pressure and the amplitude limit boundary. A preset pressure margin is established. When δP≤0 (i.e., the water pressure has reached the limit boundary), the estimated gain is reduced to zero, and the pre-adjusted incremental output is completely locked to avoid control saturation. The selection of the pressure margin must balance retaining sufficient adjustment space with preventing control saturation. In one optional scheme, the pressure margin is taken as 0.1 to 0.3 MPa.

[0105] Further, the coefficient of variation of pixel grayscale distribution within the cross-section at a specified height from the collision point of the water jet and the molten metal flow is calculated, and the density uniformity data of the cone cross-section is obtained based on the coefficient of variation. The selection of the specified height and the calculation method of the coefficient of variation are the same as in S1. The basis for determining the preset abnormal threshold is: under normal atomization conditions, the cross-sectional grayscale coefficient of variation follows an approximately normal distribution, and the density uniformity index is negatively correlated with the coefficient of variation; when the coefficient of variation deviates significantly from the normal range, it corresponds to the physical state of local nozzle blockage or spray deviation leading to significant uneven material distribution within the spray cross-section. In an optional scheme, the mean value of the cross-sectional grayscale coefficient of variation under normal operating conditions plus three times the standard deviation is taken as the upper limit of the coefficient of variation; since density uniformity is negatively correlated with the coefficient of variation, this upper limit of the coefficient of variation corresponds to the abnormal threshold of the density uniformity data of the cone cross-section. When the measured coefficient of variation exceeds this upper limit, it is determined that the density uniformity data of the cone cross-section is below the abnormal threshold.

[0106] When the uniformity of the cone's internal cross-sectional density falls below a preset abnormal threshold, an atomization anomaly is identified. The water pressure pre-adjustment increment output of the rapid prediction loop is locked, control response coefficient correction is paused, weak feedback is retained to maintain basic control, and an alarm is triggered. The weak feedback refers to reducing the feedback coefficient to 20% to 30% of the normal operating value. The lower limit of 20% ensures that the feedback loop still has basic error correction capabilities under abnormal conditions, preventing complete loss of control. The upper limit of 30% avoids excessive feedback adjustment exacerbating atomization instability in abnormal states such as nozzle blockage or uneven spray. Basic control means only the feedback loop is kept running while the rapid prediction loop output is frozen. The alarm is triggered via a pop-up window on the host computer interface, simultaneously outputting a remote notification signal. Once the uniformity of the cone's internal cross-sectional density recovers, the lockout is automatically released, and all control parameters return to their pre-abnormal state.

[0107] If the melt temperature or atomizing water pressure sensor data overflows or disconnects completely, the system will automatically lock the confidence level of the corresponding mode, switch the particle size estimation to single-mode operation mode, and trigger a sensor fault alarm simultaneously.

[0108] If the vision sensor fails completely and cannot extract any valid contours, the system degenerates into a pure process parameter feedback control mode, locking out the fast prediction loop output.

[0109] If both the melt temperature and atomizing water pressure sensors fail simultaneously, the system immediately locks the process parameter mode, retains only the visual mode, and triggers a level-two fault alarm. If both the visual mode and the process parameter mode fail completely, the system triggers a level-one fault alarm, automatically locks the water pressure to the current safe operating value, prompts for manual intervention, and prohibits all automatic adjustment command outputs.

[0110] When visual modal degradation and fogging abnormal conditions are triggered simultaneously, a lockout strategy is executed, that is, the fast prediction loop and coefficient correction function are locked at the same time, and the feedback coefficient is taken as the lower value of the two.

[0111] When a single operating condition is restored, the corresponding partial interlock is released. After all operating conditions are restored, the system returns to normal control mode. If the rate of change of the half-cone angle exceeds the rate threshold again during the ramp cancellation period, the ramp cancellation process is immediately terminated, and the system re-enters the pre-adjustment increment lock state, updating the pre-adjustment increment according to the current rate.

[0112] S4. After the water pressure adjustment is executed, the measured change in half cone angle is compared with the expected cone angle response under the corresponding pre-stored calibrated water pressure adjustment. When the deviation exceeds the preset correction threshold and the visual confidence level is not lower than the preset reliability threshold, the control response coefficient is corrected.

[0113] After water pressure adjustment is executed, a preset time window must be waited for comparison to ensure that the dynamic process of the system has basically decayed before collecting measured values, thus avoiding correction deviations caused by transient fluctuations. The time window is 3 to 5 times the response time constant of the water pressure-half-cone angle channel, that is, the time required for the half-cone angle to reach 63.2% of its steady-state value after a step change in atomized water pressure. In an optional scheme, the time window is 2 to 5 seconds.

[0114] The control response coefficient is defined as the steady-state change in powder particle size D50 caused by a unit water pressure adjustment, with dimensions in μm / MPa. Its physical meaning reflects the sensitivity of water pressure adjustment to powder particle size control under current operating conditions. In the water atomization process, an increase in atomizing water pressure leads to a decrease in powder particle size D50; therefore, the control response coefficient is negative, and its absolute value reflects the magnitude of the control sensitivity.

[0115] Since there is a stable monotonic correspondence between the half-cone angle and the particle size, this coefficient can be indirectly corrected by the deviation of the cone angle response. The premise for this correction logic is that the half-cone angle and the particle size D50 are approximately linearly negatively correlated within the working range, so that the relative deviation of the cone angle response is consistent with the relative deviation of the particle size control response. Therefore, the control response coefficient can be corrected proportionally by the relative deviation of the cone angle response.

[0116] The applicable conditions for this correction method are that the change in the half-cone angle is within the linear segment of the calibration interval, that is, Δα does not exceed ±30% of the calibration interval, and the linear approximation error of the half-cone angle-particle size mapping relationship does not exceed 5% within this range;

[0117] When the operating conditions deviate significantly from the calibration range, such as severe nozzle wear causing a significant change in the slope of the half-cone angle-particle size relationship, the linear proportional relationship between the cone angle deviation and the particle size deviation no longer holds. In this case, Kresp calibration based on the cone angle deviation should be paused, and the Kresp should be corrected directly after the next offline screening data arrives. To check whether this prerequisite is met, the system continuously monitors the consistency between the cone angle deviation and the subsequent offline screening particle size deviation in the most recent N calibrations. When the two directions are inconsistent for three consecutive times, the linear approximation is determined to have failed, and the system automatically switches to offline screening calibration mode and triggers a recalibration prompt.

[0118] The initial value of this coefficient was obtained through offline calibration. It was calculated by applying a water pressure step of known amplitude under stable operating conditions and combining it with the steady-state change in particle size obtained from offline screening.

[0119] The expected cone angle response is obtained through the following calibration procedure: under stable operating conditions, a step change of known amplitude is applied to the water pressure, and after the system response stabilizes, the steady-state change of the half cone angle is recorded. The calibration curve between the water pressure regulation and the expected cone angle response is established by using the average value of multiple repeated experiments, or by looking up a table, or by obtaining the reference response coefficient through linear fitting.

[0120] The preset calibration threshold is set based on the following: during normal calibration, the deviation between the measured cone angle response and the expected response should be within the calibration accuracy range. When the deviation significantly exceeds this range, it indicates that the current control response coefficient can no longer accurately reflect the actual operating conditions. The calibration threshold uses the trigger logic of the larger of a relative threshold and an absolute threshold. An example of a relative threshold is 10% to 20% of the expected cone angle change; exceeding this range indicates a significant deviation between the current control response coefficient and the actual operating conditions. When the water pressure adjustment is small, resulting in a very small expected cone angle change, the preset absolute threshold is used as the lower limit for calibration triggering to avoid false triggering due to measurement noise. In an optional scheme, the absolute threshold is set to 0.5° to 1.5°.

[0121] When the deviation between the measured change in the semi-cone angle and the expected change exceeds the preset correction threshold, the control response coefficient is adjusted proportionally to the deviation, as expressed in the following expression:

[0122]

[0123] in, The corrected control response coefficient; γ is the control response coefficient before correction. Its physical meaning is the steady-state change of powder D50 caused by a unit water pressure adjustment. Its dimension is μm / MPa. Since the particle size decreases as the water pressure increases, this coefficient itself is negative. γ is the preset correction step size factor, which is used to control the correction amplitude of a single correction. This represents the steady-state change in the half-cone angle measured after water pressure adjustment. The expected change in the semi-cone angle is pre-stored for calibration. In one optional scheme, the calibration step size factor is set to 0.3 to 0.7, with an upper limit of 0.7 to avoid overcalibration causing new deviations, and a lower limit of 0.3 to ensure that the calibration convergence speed can promptly track the drift caused by operating conditions such as nozzle wear.

[0124] When the measured change in the half-cone angle is less than the expected change, it indicates that the actual driving force of the water pressure on the cone angle under the current operating conditions is weaker than the calibration expectation. This may be due to nozzle wear or water temperature changes. The corresponding particle size response will also be weaker. After correction... The absolute value decreases. When the actual measurement is greater than expected, it indicates that the driving capability is stronger than expected, after correction. The absolute value of Kresp increases. After the correction is completed, the system verifies the validity of the correction result. The change in Kresp after a single correction should not exceed 20% of the current value. If it does, the correction result is not adopted, subsequent automatic corrections are suspended, and the operator is prompted to recalibrate to prevent erroneous corrections under abnormal operating conditions from causing deterioration of control performance.

[0125] After calibration, the feedback coefficient is updated synchronously to maintain a proper fit between the feedback coefficient and the control response coefficient. The control response coefficient characterizes the steady-state change in particle size caused by a unit water pressure adjustment, with dimensions in μm / MPa; the feedback coefficient characterizes the water pressure adjustment required per unit particle size deviation, with dimensions in MPa / μm. Their physical meanings are reciprocals of each other. Therefore, the fit is such that the feedback coefficient equals the preset proportional factor divided by the absolute value of the control response coefficient.

[0126]

[0127] In the formula, For feedback coefficients; To control the response coefficient, the scaling factor λ is dynamically adjusted based on the absolute value of the deviation |e| between the target particle size and the current corrected particle size estimate. When |e| is greater than a preset deviation threshold, λ takes a larger value to improve feedback sensitivity and accelerate the elimination of steady-state error. The upper limit is used to avoid overshoot oscillation caused by excessive feedback gain. When |e| is less than or equal to the deviation threshold, λ takes a smaller value to reduce feedback gain and avoid small oscillations. The lower limit ensures that the feedback loop still has basic error correction capability. In one optional scheme, the deviation threshold is 10μm to 20μm; when |e| is greater than the threshold, λ is 1.2 to 1.5, and when |e| is less than or equal to the threshold, λ is 0.8 to 1.0.

[0128] If no valid offline screening data is acquired within twice the preset maximum acquisition period, the system triggers a data timeout alarm and freezes the update permissions for the baseline offset parameters to prevent erroneous corrections based on outdated data until new valid screening data is injected. During the calibration timing window, if a new water pressure adjustment command is received, the current calibration process is terminated, and the timing is restarted based on the latest water pressure command to avoid calibration distortion caused by multiple rounds of adjustment.

[0129] Furthermore, the actual particle size data obtained from offline sieving is periodically injected into the system to correct the baseline offset of the fusion model. Since there is a transport lag between the powder and the collection tank, and the offline sieving results correspond to the atomization state at a historical moment, the transport lag duration needs to be pre-calibrated. The calibration method is as follows: a short-term step adjustment of the atomization water pressure is used to create a marker event. The step amplitude is 2% to 3% of the current water pressure, which is sufficient to form a recognizable particle size distribution change signal in the sieving data, while having negligible impact on the atomization state during normal production.

[0130] Record the moment the step is applied The particle size distribution at the collection end is continuously monitored over time to obtain the response curve of particle size deviation over time. Since the change in particle size distribution is a gradual process rather than a sudden event, the centroid moment of the response curve is used as the equivalent response moment.

[0131]

[0132] in For the first Each sampling time, This represents the change in particle size deviation relative to the baseline at that moment. The transmission hysteresis time is... This method effectively suppresses the interference of random fluctuations in the screening data on the extraction of lag time through weighted averaging.

[0133] The selection of the preset acquisition period needs to balance the sufficiency of the statistical sample and the timeliness of system drift correction. If the period is too short, the amount of powder sample collected in a single acquisition will be insufficient and the statistical representativeness of the particle size distribution will be inadequate; if the period is too long, it will be impossible to correct the systemic drift of the model in a timely manner. In one optional scheme, the preset acquisition period is set to 10 to 30 minutes. The baseline offset parameter of the fusion model is corrected based on the average deviation of multiple sets of continuous samples. To avoid baseline shifts caused by random errors in a single screening, the correction method is as follows:

[0134]

[0135] in, The corrected reference offset parameters; The reference offset parameters before correction; The average deviation of multiple consecutive samples is defined as the offline screening true value minus the corrected particle size estimate at the corresponding historical moment (a positive deviation indicates that the system estimate is too low and the baseline needs to be adjusted upward). β is the correction step size factor, which is taken from 0.3 to 0.7 to avoid overcorrection in a single correction.

[0136] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A closed-loop intelligent control method for particle size of water-atomized iron-based powder based on multimodal sensing, characterized in that, Includes the following steps: S1. Real-time acquisition of atomization cone images via a visual sensor, obtaining current melt temperature and atomization water pressure process parameters, and extracting the atomization cone half-cone angle from the atomization cone image. The visual sensor employs a short-wave infrared thermal imager, utilizing the difference in radiance between high-temperature metal droplets and the background of steam and water mist in the short-wave infrared band to extract the outline of the metal droplets; S1 includes: Background subtraction and filtering are performed on the acquired fog cone image, and the fog cone contour boundary is extracted based on the processed image; Taking the point where the water jet collides with the molten metal as the vertex, draw tangents along the contours of both sides of the cone, and calculate half of the angle between the two tangents to obtain the half cone angle; Calculate the diameter of the cone section at a specified height from the point of collision between the water jet and the molten metal flow to obtain the radial width data of the spray. Calculate the coefficient of variation of pixel grayscale distribution within a cross section at a specified height from the collision point of the water jet and the molten metal flow to obtain the density uniformity data of the cone cross section; The semi-cone angle, spray radial width data, and density uniformity data are used as feature inputs for the visual modality; S2. Determine the visual confidence level based on the quality of the atomized cone image, and determine the process parameter modal confidence level based on the melt temperature and atomized water pressure process parameters. Determine the fusion weight of the visual modal and the process parameter modal based on the visual confidence level and the process parameter modal confidence level. Weight the particle size estimation result based on the half cone angle and the particle size estimation result based on the melt temperature and atomized water pressure to obtain the current particle size estimate. When the visual confidence level is lower than the preset drop trigger threshold, the visual modal fusion weight is reduced to the preset lower limit, the particle size estimation degenerates into a pure process parameter driven mode, and the water pressure pre-adjustment increment output of the fast prediction loop is simultaneously locked. When the visual confidence level rises above the preset recovery threshold, the fusion weight gradually recovers at the preset rate, and the water pressure pre-adjustment incremental output of the rapid prediction loop is simultaneously unlocked; the descent trigger threshold is lower than the rise recovery threshold, and the two constitute a hysteresis band. S3. When the rate of change of the half-cone angle exceeds the preset rate threshold, calculate the water pressure pre-adjustment increment according to the preset estimated gain and output it. The pre-adjustment increment is locked until the rate of change of the half-cone angle falls back below the rate threshold. When the particle size estimate deviates from the target particle size, calculate the water pressure correction amount according to the preset feedback coefficient. Superimpose the water pressure pre-adjustment increment and the water pressure correction amount to generate a water pressure adjustment command and output it for execution. Calculate the coefficient of variation of the pixel grayscale distribution in the cross section at a specified height from the collision point of the water jet and the molten metal flow. Based on the coefficient of variation, obtain the density uniformity data of the cross section inside the cone. When the density uniformity data of the cross section inside the cone is lower than the preset abnormal threshold, it is determined to be an abnormal atomization condition. Lock the water pressure pre-adjustment increment output of the fast prediction loop, suspend the control response coefficient correction, retain the weak feedback to maintain basic control and trigger an alarm. The locking will be automatically released once the uniformity data of the density of the cone's internal cross section is restored. S4. After the water pressure adjustment is executed, the measured change in half cone angle is compared with the expected cone angle response under the corresponding pre-stored calibrated water pressure adjustment. When the deviation exceeds the preset correction threshold and the visual confidence is not lower than the preset reliability threshold, the control response coefficient is corrected. The actual particle size data obtained from offline screening is periodically injected into the system to correct the reference offset of the fusion model.

2. The closed-loop intelligent control method for particle size of water-atomized iron-based powder based on multimodal sensing according to claim 1, characterized in that, The S1 further includes: interpolating and resampling the melt temperature and atomizing water pressure process parameters according to the time of the visual frame, so that each visual image corresponds to a set of time-synchronized process parameter values.

3. The closed-loop intelligent control method for particle size of water-atomized iron-based powder based on multimodal sensing according to claim 1, characterized in that, In S2, the visual confidence is calculated by combining the contrast of the current frame image, the edge gradient intensity, and the integrity of the fog cone contour. The modal confidence level of the process parameters is calculated from the coefficient of variation of melt temperature and atomizing water pressure within the recent sliding window.

4. The closed-loop intelligent control method for particle size of water-atomized iron-based powder based on multimodal sensing according to claim 1, characterized in that, S3 further includes: The upper and lower limits and the rate of change of the water pressure regulation command are set; when the difference between the water pressure and the upper or lower limit is less than the preset pressure margin, the estimated gain is reduced accordingly according to the ratio of the difference to the pressure margin.

5. The closed-loop intelligent control method for particle size of water-atomized iron-based powder based on multimodal sensing according to claim 1, characterized in that, In step S4, when the measured change in the half-cone angle is less than the expected change, the absolute value of the control response coefficient is decreased; when the measured change in the half-cone angle is greater than the expected change, the absolute value of the control response coefficient is increased. After calibration, the feedback coefficients are updated synchronously to ensure that the feedback coefficients and control response coefficients are in a suitable relationship.

Images