Method and system for regulating nickel-based superalloy surface molten pool flow and solidification

Abstract

The application provides a kind of nickel-based superalloy surface molten pool flow and solidification regulation method and system, the method comprises: real-time acquisition molten pool image by imaging device, extract width, length and centroid lateral deviation, construct three-dimensional molten pool morphology feature vector, and difference with target vector to obtain morphology error vector, adopt composite nonlinear control law, error vector is mapped into virtual control instruction vector by dimension conversion matrix, switch function is switched between non-singular fast terminal sliding mode and continuous high-order polynomial control according to error norm, output corresponding mode of operation discrete state signal, filter virtual instruction using finite time convergence instruction filter, parameters are scheduled by discrete state signal;Combined with coupled error compensation system to generate compensation vector, sum filtered instruction and compensation signal to obtain three-dimensional control quantity of laser power, scanning speed, lateral path offset, molten pool regulation is realized.

Description

Technical Field

[0001] This application belongs to the field of regulation, and in particular relates to a method and system for regulating the flow and solidification of the molten pool on the surface of a nickel-based superalloy. Background Technology

[0002] Nickel-based superalloys play an indispensable role in cutting-edge manufacturing fields such as aerospace engines, gas turbines, and nuclear reactors due to their excellent high-temperature strength, oxidation resistance, and corrosion resistance. Advanced processing technologies such as laser additive manufacturing, laser cladding, and laser polishing, through the interaction of high-energy laser beams with alloy materials, achieve integrated forming of complex components and high-performance surface preparation. The laser processing process is a nonlinear, strongly coupled system, susceptible to multiple random disturbances such as laser power fluctuations, powder feed rate variations, scanning path curvature, and internal heat accumulation effects within the workpiece. This can easily lead to unstable molten pool morphology, inducing defects such as porosity, cracks, and lack of fusion, thus limiting the reliability and lifespan of nickel-based superalloy components. By real-time monitoring of the molten pool image, extracting key morphological features, and planning controllers to adjust process parameters such as laser power and scanning speed, morphological errors can be suppressed.

[0003] However, traditional linear control methods such as PID controllers cannot handle the strong nonlinear characteristics of molten pool systems. They often respond slowly to disturbances and may overshoot near steady state, failing to balance speed and stability. To protect actuators such as lasers and galvanometers and prevent saturation, the output commands of the control system typically need to be smoothed by command filters. However, while traditional command filters remove high-frequency noise, they inevitably introduce time delays and phase lags, reducing the response bandwidth and tracking accuracy of the control system. When the control system needs to switch between a strong response mode to handle sudden disturbances and a steady-state regulation mode, static filtering and compensation strategies cannot match different performance requirements, becoming a technical obstacle restricting further improvement in molten pool control performance. Summary of the Invention

[0004] To address the problem that existing technologies cannot simultaneously achieve both speed and stability, inevitably resulting in time delays and phase lag, and failing to meet different performance requirements.

[0005] In the first aspect, the present invention provides the following technical solution: a method for controlling the flow and solidification of a molten pool on the surface of a nickel-based superalloy, comprising the following steps:

[0006] Image information of the molten pool is acquired in real time, and a three-dimensional molten pool morphology feature vector is constructed. The difference between the real-time extracted three-dimensional molten pool morphology feature vector and a preset target morphology feature vector is calculated to obtain a three-dimensional molten pool morphology error vector.

[0007] A composite nonlinear control law is constructed, which maps and converts the three-dimensional molten pool morphology error vector into a three-dimensional virtual control command vector through a mapping relationship including a dimension transformation matrix. The control law has a built-in switching function, which switches between a non-singular fast terminal sliding mode control mode and a continuous high-order polynomial control mode based on whether the norm of the three-dimensional molten pool morphology error vector is greater than a preset switching threshold, and outputs a discrete state signal according to the current control mode.

[0008] The three-dimensional virtual control command vector is filtered by a command filter, the parameters of which are scheduled by the discrete state signal. An error compensation system coupled with the control law mode is established, generating a compensation signal vector based on the error generated in the current filtering process and the prediction of future filtering errors. The mathematical model used to predict future filtering errors and the prediction time domain length are selected or adjusted by the discrete state signal. The three-dimensional control quantity is obtained based on the filtered command vector output after processing by the command filter and the compensation signal vector generated by the error compensation system.

[0009] Optionally, the real-time acquisition of image information of the molten pool and the construction of a three-dimensional molten pool morphology feature vector include:

[0010] A camera is used to acquire grayscale images of the molten pool with a resolution of 1280×1024 pixels at a rate of no less than 150 frames per second. The acquired grayscale images are then subjected to Gaussian filtering with a 5×5 pixel kernel for noise reduction and image binarization based on Otsu's method to obtain a binary mask of the molten pool region.

[0011] Extract the maximum connected region of the binary mask as the molten pool profile, calculate the width of the minimum bounding rectangle of the molten pool profile as the molten pool width, calculate the height of the minimum bounding rectangle of the molten pool profile as the molten pool length, and calculate the difference between the coordinates of the centroid of the profile perpendicular to the predetermined laser travel direction and the coordinates of the center line of the travel direction as the lateral deviation of the molten pool centroid.

[0012] The values ​​of width, length, and lateral deviation are arranged in a predetermined order to form a 3×1 column vector, which serves as the feature vector of the three-dimensional molten pool morphology.

[0013] Optionally, the control law incorporates a switching function. This function switches between a non-singular fast terminal sliding mode control mode and a continuous high-order polynomial control mode based on whether the norm of the three-dimensional melt pool morphology error vector is greater than a preset switching threshold. Based on the current control mode, it outputs a discrete state signal, including:

[0014] Calculate the L2 norm of the three-dimensional molten pool morphology error vector and set a fixed switching threshold;

[0015] When the L2 norm is not less than the switching threshold, the control law is switched to non-singular fast terminal sliding mode control mode, and a discrete state signal representing the strong response mode is output.

[0016] When the L2 norm is less than the switching threshold, the control law is switched to a continuous high-order polynomial control mode, and a discrete state signal representing the fine-tuning smoothing mode is output.

[0017] Optionally, the parameters of the instruction filter are scheduled by the discrete state signal, including:

[0018] When the discrete-state signal is in a strong response mode, the boundary layer thickness parameter matrix of the saturation function in the command filter is set as a diagonal matrix, with the diagonal elements being the product of k1 and the channel scaling factor. Simultaneously, the gain coefficient matrix is... , Set as a diagonal matrix, with diagonal elements being the product of k2 and the channel scaling factor;

[0019] When the discrete state signal is in fine-tuned smoothing mode, the diagonal elements of the boundary layer thickness parameter matrix are set to the product of k3 and the channel scaling factor, and the diagonal elements of the two gain coefficient matrices are set to the product of k4 and the channel scaling factor.

[0020] Optionally, the mathematical model used to predict future filtering errors and the prediction time domain length are selected or adjusted by the discrete state signal, including:

[0021] When the discrete state signal is in a strong response mode, a lightweight recurrent neural network based on a gated recurrent unit is selected to predict the future filtering error, and the prediction time domain length is set to the next 4 control cycles.

[0022] When the discrete state signal is in a fine-tuned smoothing mode, a second-order linear autoregressive model is selected to predict future filtering errors, and the prediction time domain length is set to the next two control cycles.

[0023] Optionally, obtaining the three-dimensional control quantity based on the filtered command vector output after processing by the command filter and the compensation signal vector generated by the error compensation system includes:

[0024] In each control cycle, the filtered virtual control command vector output by the command filter is added to the compensation signal vector output by the error compensation system to obtain the three-dimensional control quantity of the cycle. The three-dimensional control quantity includes laser power, scanning speed and lateral path offset.

[0025] In a second aspect, the present invention provides the following technical solution: a system for controlling the flow and solidification of a molten pool on the surface of a nickel-based superalloy, comprising the following modules:

[0026] The construction module is used to acquire image information of the molten pool in real time and construct a three-dimensional molten pool morphology feature vector; the difference between the real-time extracted three-dimensional molten pool morphology feature vector and a preset target morphology feature vector is calculated to obtain a three-dimensional molten pool morphology error vector.

[0027] The output module is used to construct a composite nonlinear control law. Through a mapping relationship including a dimension transformation matrix, it maps and converts the three-dimensional molten pool morphology error vector into a three-dimensional virtual control command vector. The control law has a built-in switching function. The function switches between a non-singular fast terminal sliding mode control mode and a continuous high-order polynomial control mode based on whether the norm of the three-dimensional molten pool morphology error vector is greater than a preset switching threshold. Based on the current control mode, it outputs a discrete state signal.

[0028] A generation module is used to filter the three-dimensional virtual control command vector using a command filter, the parameters of which are scheduled by the discrete state signal; an error compensation system coupled with the control law mode is established, which generates a compensation signal vector based on the error generated in the current filtering process and the prediction of future filtering errors, wherein the mathematical model used to predict future filtering errors and the prediction time domain length are selected or adjusted by the discrete state signal; and a three-dimensional control quantity is obtained based on the filtered command vector output after processing by the command filter and the compensation signal vector generated by the error compensation system.

[0029] Preferably, the real-time acquisition of image information of the molten pool and the construction of a three-dimensional molten pool morphology feature vector include:

[0030] A camera is used to acquire grayscale images of the molten pool with a resolution of 1280×1024 pixels at a rate of no less than 150 frames per second. The acquired grayscale images are then subjected to Gaussian filtering with a 5×5 pixel kernel for noise reduction and image binarization based on Otsu's method to obtain a binary mask of the molten pool region.

[0031] Extract the maximum connected region of the binary mask as the molten pool profile, calculate the width of the minimum bounding rectangle of the molten pool profile as the molten pool width, calculate the height of the minimum bounding rectangle of the molten pool profile as the molten pool length, and calculate the difference between the coordinates of the centroid of the profile perpendicular to the predetermined laser travel direction and the coordinates of the center line of the travel direction as the lateral deviation of the molten pool centroid.

[0032] The values ​​of width, length, and lateral deviation are arranged in a predetermined order to form a 3×1 column vector, which serves as the feature vector of the three-dimensional molten pool morphology.

[0033] Preferably, the control law incorporates a switching function. This function switches between a non-singular fast terminal sliding mode control mode and a continuous high-order polynomial control mode based on whether the norm of the three-dimensional melt pool morphology error vector is greater than a preset switching threshold. Based on the current control mode, it outputs a discrete state signal, including:

[0034] Calculate the L2 norm of the three-dimensional molten pool morphology error vector and set a fixed switching threshold;

[0035] When the L2 norm is not less than the switching threshold, the control law is switched to non-singular fast terminal sliding mode control mode, and a discrete state signal representing the strong response mode is output.

[0036] When the L2 norm is less than the switching threshold, the control law is switched to a continuous high-order polynomial control mode, and a discrete state signal representing the fine-tuning smoothing mode is output.

[0037] Preferably, the parameters of the instruction filter are scheduled by the discrete state signal, including:

[0038] When the discrete-state signal is in a strong response mode, the boundary layer thickness parameter matrix of the saturation function in the command filter is set as a diagonal matrix, with the diagonal elements being the product of k1 and the channel scaling factor. Simultaneously, the gain coefficient matrix is... , Set as a diagonal matrix, with diagonal elements being the product of k2 and the channel scaling factor;

[0039] When the discrete state signal is in fine-tuned smoothing mode, the diagonal elements of the boundary layer thickness parameter matrix are set to the product of k3 and the channel scaling factor, and the diagonal elements of the two gain coefficient matrices are set to the product of k4 and the channel scaling factor.

[0040] Preferably, the mathematical model used to predict future filtering errors and the prediction time domain length are selected or adjusted by the discrete state signal, including:

[0041] When the discrete state signal is in a strong response mode, a lightweight recurrent neural network based on a gated recurrent unit is selected to predict the future filtering error, and the prediction time domain length is set to the next 4 control cycles.

[0042] When the discrete state signal is in a fine-tuned smoothing mode, a second-order linear autoregressive model is selected to predict future filtering errors, and the prediction time domain length is set to the next two control cycles.

[0043] Preferably, obtaining the three-dimensional control quantity based on the filtered command vector output after processing by the command filter and the compensation signal vector generated by the error compensation system includes:

[0044] In each control cycle, the filtered virtual control command vector output by the command filter is added to the compensation signal vector output by the error compensation system to obtain the three-dimensional control quantity of the cycle. The three-dimensional control quantity includes laser power, scanning speed and lateral path offset.

[0045] This invention provides a method for controlling the molten pool on the surface of nickel-based superalloys. By designing a composite nonlinear control law that can switch between a strong response mode and a fine-tuning smoothing mode, it balances the system's ability to rapidly correct large-scale morphological deviations with its ability to finely smooth small-scale fluctuations. By constructing a finite-time convergent command filter and error compensation system coupled with the control law mode, it suppresses control command chattering, ensures smooth transition of control quantities, and compensates for errors in the filtering process, eliminating lag and deviation caused by the command filtering stage and improving the system's tracking accuracy. By treating the width, length, and lateral deviation of the molten pool as unified control objects, it achieves decoupling and coordinated control of the multi-dimensional morphological characteristics of the molten pool. This invention realizes closed-loop control of the molten pool morphology on the surface of nickel-based superalloys, ensuring the quality and consistency of machined parts under complex working conditions. Attached Figure Description

[0046] Figure 1 This is a schematic diagram of the hardware device;

[0047] Figure 2 This is a flowchart of a method for controlling the flow and solidification of the molten pool on the surface of a nickel-based superalloy. Detailed Implementation

[0048] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this specification. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this specification as detailed in the appended claims.

[0049] It should be understood that the terms “comprising” and “having”, and any variations thereof, in the embodiments of this specification are intended to cover but not exclude inclusion. For example, a product or device that includes a series of components is not necessarily limited to those components that are explicitly listed, but may include other components that are not explicitly listed or that are inherent to such product or device.

[0050] Example 1

[0051] Hardware devices such as Figure 1 As shown, it mainly consists of an infrared thermal imager, a high frame rate camera, a control computer, a laser cladding head, a laser generator, a powder feeder, a worktable, and a substrate. In Embodiment 1 of this invention, a method for controlling the flow and solidification of the molten pool on the surface of a nickel-based superalloy is described, as follows: Figure 2 As shown, it includes the following steps:

[0052] S1. Acquire image information of the molten pool in real time and construct a three-dimensional molten pool morphology feature vector; subtract the three-dimensional molten pool morphology feature vector extracted in real time from a preset target morphology feature vector to obtain a three-dimensional molten pool morphology error vector.

[0053] A near-infrared CMOS camera with a frame rate of at least 150Hz, configured coaxially or off-axis, is used as the imaging device. Hardware synchronization is achieved through a camera data acquisition card and a motion control system to ensure that the image frames correspond to the position of the laser scanning head. For each 8-bit grayscale image acquired in real time, a 5x5 kernel mean filtering algorithm is used for initial noise reduction. A threshold segmentation algorithm with Gaussian weighted average as the core is applied to separate the molten pool region from the surrounding matrix background, generating a binarized image. The findContours function from the open-source computer vision library OpenCV is executed on the binarized image to search for all contours, and the unique contour representing the molten pool is selected based on the principle of maximizing the contour area. The minAreaRect function is called to fit the contour with a minimum bounding rectangle, where the long side and short side of the rectangle are defined as the length and width of the molten pool, respectively. The moments function is called to calculate the image moments of the contour, and the centroid coordinates of the molten pool are solved by the ratio of the first moment to the zero moment. Substituting the centroid coordinates into the pre-calibrated transformation matrix between the camera coordinate system and the workpiece coordinate system, the position of the centroid in the workpiece coordinate system is calculated. Then, using the formula for calculating the distance from a point to a line, the Euclidean distance between the centroid position and the centerline of the predetermined machining path is obtained; this is the lateral deviation. The obtained width, length, and lateral deviation floating-point values ​​are sequentially stored into a one-dimensional array to form a real-time three-dimensional molten pool morphology feature vector. This vector is then subtracted element-wise from a preset target morphology feature vector containing the target width, target length, and zero deviation to generate a three-dimensional molten pool morphology error vector.

[0054] In an alternative embodiment, the real-time acquisition of image information of the molten pool and the construction of a three-dimensional molten pool morphology feature vector include:

[0055] A camera is used to acquire grayscale images of the molten pool with a resolution of 1280×1024 pixels at a rate of no less than 150 frames per second. The acquired grayscale images are then subjected to Gaussian filtering with a 5×5 pixel kernel for noise reduction and image binarization based on Otsu's method to obtain a binary mask of the molten pool region.

[0056] Extract the maximum connected region of the binary mask as the molten pool profile, calculate the width of the minimum bounding rectangle of the molten pool profile as the molten pool width, calculate the height of the minimum bounding rectangle of the molten pool profile as the molten pool length, and calculate the difference between the coordinates of the centroid of the profile perpendicular to the predetermined laser travel direction and the coordinates of the center line of the travel direction as the lateral deviation of the molten pool centroid.

[0057] The values ​​of width, length, and lateral deviation are arranged in a predetermined order to form a 3×1 column vector, which serves as the feature vector of the three-dimensional molten pool morphology.

[0058] A coaxial vision camera was used to acquire images at a rate of 200 frames per second, ensuring real-time monitoring of the molten pool process. Each frame had a resolution of 1280×1024 pixels. The acquired raw grayscale images were processed by a 5×5 pixel Gaussian filter with a standard deviation σ of 1.5 to smooth high-frequency noise from the CMOS sensor input while avoiding excessive blurring of molten pool edge details. Otsu's method was used to perform global threshold binarization on the denoised images. This algorithm automatically finds the optimal threshold based on the image's grayscale histogram, separating the bright foreground molten pool area from the darker substrate background to generate a binary mask. A connected component analysis algorithm was then applied to this binary mask to identify the largest connected component as the target molten pool area, filtering out isolated noise points caused by factors such as metal spatter. Based on the extracted molten pool contour, the minimum area bounding rectangle was calculated, with the shorter side defined as the molten pool width W and the longer side as the molten pool length L. Simultaneously, the geometric centroid coordinates of this contour were calculated. Assuming the centerline of the preset processing path corresponds to the center column of the image, that is =640, then the lateral deviation of the centroid Through formula The calculation yields a value where S is a pre-calibrated pixel-to-physical-size conversion factor, for example, 0.01 mm / pixel. The three scalar values ​​W, L, and... All units are in mm, according to The sequence is used to construct a 3×1 three-dimensional molten pool morphology feature vector Y. For example, the measurement result at a certain moment might be... .

[0059] S2, Construct a composite nonlinear control law. Through a mapping relationship including a dimension transformation matrix, the three-dimensional molten pool morphology error vector is mapped and converted into a three-dimensional virtual control command vector. The control law has a built-in switching function. The function switches between a non-singular fast terminal sliding mode control mode and a continuous high-order polynomial control mode based on whether the norm of the three-dimensional molten pool morphology error vector is greater than a preset switching threshold. The function also outputs a discrete state signal based on the current control mode.

[0060] A 3x3 diagonal matrix is ​​defined as the dimension transformation matrix, with diagonal elements calibrated based on physical experimental data. This matrix is ​​used to convert geometric errors in millimeters into dimensionless control error signals. The L2 norm of the three-dimensional molten pool morphology error vector is calculated, which is the square root of the sum of the squares of the vector's elements. This L2 norm is then compared to a preset floating-point switching threshold. If the norm is greater than the threshold, the non-singular fast terminal sliding mode control mode is activated. In this mode, the discrete state signal output is 1, and the control law is used to calculate each component of the three-dimensional virtual control command vector in the following form. : ,in Let i be the i-th component of the error vector. , , p and q are preset sliding mode control parameters. Let be the error integral term, and 'sign' be the sign function. If the norm is not greater than this threshold, the system switches to continuous high-order polynomial control mode. In this mode, the discrete-state signal output is 0, and the control law is used to calculate each component of the virtual control command vector in the following form. : ,in , , The coefficients are preset polynomials, which provide a smooth and continuous control output when the error is close to zero.

[0061] In some embodiments, the control law incorporates a switching function that switches between a non-singular fast terminal sliding mode control mode and a continuous high-order polynomial control mode based on whether the norm of the three-dimensional melt pool morphology error vector is greater than a preset switching threshold, and outputs a discrete state signal according to the current control mode, including:

[0062] Calculate the L2 norm of the three-dimensional molten pool morphology error vector and set a fixed switching threshold;

[0063] When the L2 norm is not less than the switching threshold, the control law is switched to non-singular fast terminal sliding mode control mode, and a discrete state signal representing the strong response mode is output.

[0064] When the L2 norm is less than the switching threshold, the control law is switched to a continuous high-order polynomial control mode, and a discrete state signal representing the fine-tuning smoothing mode is output.

[0065] Define a target morphological feature vector ,For example The unit is mm. The real-time measured feature vector Y is compared with the target vector. The difference is used to obtain the three-dimensional molten pool morphology error vector. Calculate the L2 norm of the error vector, i.e., the Euclidean norm: Set a fixed switching threshold ε, the value of which needs to balance response speed and steady-state accuracy. A reasonable value is 5%-10% of the target vector norm. For example, if... If the value is approximately 2.33mm, then ε can be set to 0.15mm. Within each control cycle, if the calculated value... The value mm indicates that the system state deviates significantly from the target. At this point, the control law is switched to Non-Singular Fast Terminal Sliding Mode Control (NFTSMC). This control mode ensures that the error converges rapidly within a finite time, and the system outputs a discrete state signal. =1 indicates that the system has entered a strong response mode; conversely, if If the value is less than 0.15mm, it indicates that the system has entered the target's neighborhood. At this point, to avoid the chattering problem inherent in sliding mode control and to pursue higher steady-state accuracy, the control law is seamlessly switched to a preset continuous high-order polynomial control mode. In this mode, the control output is smoother, and the system outputs discrete state signals. =0 indicates that the system has entered the fine-tuning and smoothing mode.

[0066] S3, the three-dimensional virtual control command vector is filtered by a command filter, the parameters of which are scheduled by the discrete state signal; an error compensation system coupled with the control law mode is established, and a compensation signal vector is generated based on the error generated in the current filtering process and the prediction of future filtering errors, wherein the mathematical model used to predict future filtering errors and the prediction time domain length are selected or adjusted by the discrete state signal; the three-dimensional control quantity is obtained based on the filtered command vector output after processing by the command filter and the compensation signal vector generated by the error compensation system.

[0067] A second-order command filter based on high-order sliding mode observer theory uses a saturation function in its differential equation to eliminate chattering, and its state equation is as follows: ,in This is the filtered instruction. For the derivative, The input virtual control command vector; The convergence rate gain matrix is... The nonlinear damping gain matrix is... Saturation function The boundary layer thickness diagonal matrix; the saturation function is defined as Exponential parameters When the discrete state signal is 1, i.e., in strong response mode, a set of diagonal matrices is selected. and smaller boundary layer thickness To achieve rapid convergence. When the discrete state signal is 0, i.e., in fine-tuning smoothing mode, it switches to another set of larger diagonal matrices. and a large boundary layer thickness This results in a smoother filtering effect. The error compensation system calculates the current filtering error vector. For predicting future filtering errors, when the discrete state signal is 0, a low-computation single-step Euler forward difference method is used. Based on the current state of the filter and the virtual command for the next control cycle, the filtering error within a short time window is predicted. When the discrete state signal is 1, the more computationally accurate fourth-order Runge-Kutta method is invoked to perform multi-step numerical integration on the filter differential equation, predicting the filtering error trajectory within a longer time window. The compensation signal vector is generated through a proportional-integral controller, with the input being the weighted sum of the current filtering error and the predicted filtering error.

[0068] The filter command vector output by the filter The three-dimensional control vector is obtained by performing vector addition with the compensation signal vector. The first component of the vector is multiplied by a power gain coefficient and added to a nominal laser power value to obtain the laser power command. The second component is multiplied by a velocity gain coefficient and added to a nominal scanning velocity value to obtain the scanning velocity command. The third component is multiplied by a displacement gain coefficient and used as a lateral path offset command, which is updated in real time to the motion controller of the laser scanning galvanometer.

[0069] In other embodiments, the parameters of the instruction filter are scheduled by the discrete state signal, including:

[0070] When the discrete-state signal is in a strong response mode, the boundary layer thickness parameter matrix of the saturation function in the command filter is set as a diagonal matrix, with the diagonal elements being the product of k1 and the channel scaling factor. Simultaneously, the gain coefficient matrix is... , Set as a diagonal matrix, with diagonal elements being the product of k2 and the channel scaling factor;

[0071] When the discrete state signal is in fine-tuned smoothing mode, the diagonal elements of the boundary layer thickness parameter matrix are set to the product of k3 and the channel scaling factor, and the diagonal elements of the two gain coefficient matrices are set to the product of k4 and the channel scaling factor.

[0072] Preferably, k1 is 0.2, k2 is 1.5, k3 is 2.0, and k4 is 10. To handle the different meanings and orders of magnitude of the components in the virtual control command vector, a channel scaling diagonal matrix is ​​input. The diagonal elements are tuned according to the response characteristics of each control channel; for example, the laser power, scanning speed, and lateral offset can be set. =100, =500, =1. The three key parameter matrices within the instruction filter, boundary layer thickness... Gain matrix and All based on The value is adjusted. When In the strong response mode (=1), the goal is to quickly track virtual control commands. In this mode, the boundary layer thickness parameter matrix... Set it to 0.2×C=diag(20,100,0.2); at the same time, set the gain coefficient matrix... and Setting it to 1.5×C=diag(150,750,1.5), the moderate gain ensures fast and stable convergence. When In the fine-tuning smoothing mode with =0, the goal shifts to suppressing jitter and ensuring smooth output. At this point, the boundary layer thickness parameter matrix... Setting it to 2.0×C=diag(200,1000,2.0), the thickened boundary layer makes the saturation function approximate a linear function when the error is small, thus obtaining a smooth filtered output; at the same time, the gain coefficient matrix... and Setting it to 10.0×C=diag(1000,5000,10.0) provides a higher gain, which offers stronger error suppression in the quasi-linear region and ensures high steady-state accuracy.

[0073] In an alternative embodiment, the mathematical model for predicting future filtering errors and the prediction time domain length are selected or adjusted by the discrete state signal, including:

[0074] When the discrete state signal is in a strong response mode, a lightweight recurrent neural network based on a gated recurrent unit is selected to predict the future filtering error, and the prediction time domain length is set to the next 4 control cycles.

[0075] When the discrete state signal is in a fine-tuned smoothing mode, a second-order linear autoregressive model is selected to predict future filtering errors, and the prediction time domain length is set to the next two control cycles.

[0076] The prediction module in the error compensation system also consists of discrete state signals. Configure it. When In the strong response mode (=1), the system is in a large transient process and exhibits strong nonlinear behavior. Therefore, a GRU network with stronger expressive power is selected to predict the filtering error. The GRU network is a lightweight recurrent neural network specifically designed for processing and predicting time series data, and is pre-trained offline. The network structure includes an input layer, a GRU hidden layer, and a fully connected output layer. The network input is a time series matrix. Where N is the historical time step, for example N=10, and 6 represents the input feature dimension at each time step; this input feature vector From the three-dimensional filtering error vector and three-dimensional virtual control command vector Series connection, i.e. The GRU hidden layer contains 32 units, each with update and reset gates to learn intermediate dependencies in the sequence. The activation functions are the sigmoid function and the hyperbolic tangent function. The fully connected output layer maps the output of the GRU hidden layer at the last time step to the prediction result, outputting a prediction sequence matrix. ,in To predict the time domain length, for example =4, 3 represents the predicted three-dimensional filtering error vector at each future time step. ,Right now At this point, the time domain length will be predicted. The control cycle is set to 4 cycles. If the control frequency is 200Hz, it corresponds to 20 milliseconds. The longer prediction time domain helps to compensate for the phase lag of the filter input during the transient process.

[0077] when When the smoothing mode is fine-tuned to 0, the system fluctuates within a small range around the target point and can be approximated as a linear time-invariant system. Therefore, to reduce the computational burden, we switch to a second-order linear autoregressive AR model. This model independently models each component of the filtering error, in the form of... , where the coefficient , Real-time identification can be performed using online recursive least squares (RLS), or a fixed value obtained from offline identification can be used. In this case, the prediction time domain length... The control cycle is shortened to two cycles, corresponding to 10ms, because the filtering error is small and changes gradually in the linear region, and short time-domain prediction is accurate enough, while avoiding the risk of error accumulation caused by long time-domain prediction.

[0078] In some embodiments, obtaining the three-dimensional control quantity based on the filtered command vector output after processing by the command filter and the compensation signal vector generated by the error compensation system includes:

[0079] In each control cycle, the filtered virtual control command vector output by the command filter is added to the compensation signal vector output by the error compensation system to obtain the three-dimensional control quantity of the cycle. The three-dimensional control quantity includes laser power, scanning speed and lateral path offset.

[0080] The command filter outputs a smoothed virtual control command vector. At the same time, the error compensation system adjusts according to the current... The selected model and its corresponding prediction time domain generate a compensation signal vector. The active cancellation filter will generate errors, and one way to implement the compensation signal is to predict future errors. The prediction errors of each step are weighted and summed. The filter command and compensation signal are then combined using vector addition to obtain the control command vector. . Vector The three components are mapped to the executor. For example, Set to the output power of the laser, such as =1250, then the laser power is set to 1250 watts. This value must be within the actuator's allowable range, such as 800-2000 watts. Set to the galvanometer scanning speed, such as =950, then the scanning speed is set to 950 mm / s, and this value must be within the galvanometer performance range, such as 500-1500 mm / s. This becomes the offset of the lateral path, superimposed on the predetermined path. For example, if the y-coordinate of the predetermined path at the current time is... =25.4mm, and =-0.12mm, then the instantaneous target y-coordinate command of the galvanometer is: =25.28mm, and this offset is usually limited to a small range, such as ±0.5mm.

[0081] Example 2

[0082] In Embodiment 2 of the present invention, a system for controlling the flow and solidification of the molten pool on the surface of a nickel-based superalloy is provided, comprising the following modules:

[0083] The construction module is used to acquire image information of the molten pool in real time and construct a three-dimensional molten pool morphology feature vector; the difference between the real-time extracted three-dimensional molten pool morphology feature vector and a preset target morphology feature vector is calculated to obtain a three-dimensional molten pool morphology error vector.

[0084] The output module is used to construct a composite nonlinear control law. Through a mapping relationship including a dimension transformation matrix, it maps and converts the three-dimensional molten pool morphology error vector into a three-dimensional virtual control command vector. The control law has a built-in switching function. The function switches between a non-singular fast terminal sliding mode control mode and a continuous high-order polynomial control mode based on whether the norm of the three-dimensional molten pool morphology error vector is greater than a preset switching threshold. Based on the current control mode, it outputs a discrete state signal.

[0085] A generation module is used to filter the three-dimensional virtual control command vector using a command filter, the parameters of which are scheduled by the discrete state signal; an error compensation system coupled with the control law mode is established, which generates a compensation signal vector based on the error generated in the current filtering process and the prediction of future filtering errors, wherein the mathematical model used to predict future filtering errors and the prediction time domain length are selected or adjusted by the discrete state signal; and a three-dimensional control quantity is obtained based on the filtered command vector output after processing by the command filter and the compensation signal vector generated by the error compensation system.

[0086] It should be understood that in the foregoing description of the embodiments in this specification, various features are combined in a single embodiment, drawing, or description for the purpose of simplifying the description and to aid in understanding a feature. However, this does not mean that the combination of these features is necessary, and those skilled in the art, upon reading this specification, may readily identify some of the devices as separate embodiments. That is, the embodiments in this specification can also be understood as an integration of multiple secondary embodiments. And the content of each secondary embodiment is valid even if it contains fewer than all the features of a single foregoing disclosed embodiment.

Claims

1. A method for controlling the flow and solidification of the molten pool on the surface of a nickel-based superalloy, characterized in that, Includes the following steps: Image information of the molten pool is acquired in real time, and a three-dimensional molten pool morphology feature vector is constructed. The difference between the real-time extracted three-dimensional molten pool morphology feature vector and a preset target morphology feature vector is calculated to obtain a three-dimensional molten pool morphology error vector. A composite nonlinear control law is constructed, which maps and converts the three-dimensional molten pool morphology error vector into a three-dimensional virtual control command vector through a mapping relationship including a dimension transformation matrix. The control law has a built-in switching function, which switches between a non-singular fast terminal sliding mode control mode and a continuous high-order polynomial control mode based on whether the norm of the three-dimensional molten pool morphology error vector is greater than a preset switching threshold, and outputs a discrete state signal according to the current control mode. The three-dimensional virtual control command vector is filtered by a command filter, the parameters of which are scheduled by the discrete state signal. An error compensation system coupled with the control law mode is established, generating a compensation signal vector based on the error generated in the current filtering process and the prediction of future filtering errors. The mathematical model used to predict future filtering errors and the prediction time domain length are selected or adjusted by the discrete state signal. The three-dimensional control quantity is obtained based on the filtered command vector output after processing by the command filter and the compensation signal vector generated by the error compensation system.

2. The method according to claim 1, characterized in that, The real-time acquisition of image information of the molten pool and the construction of a three-dimensional molten pool morphology feature vector include: A camera is used to acquire grayscale images of the molten pool with a resolution of 1280×1024 pixels at a rate of no less than 150 frames per second. The acquired grayscale images are then subjected to Gaussian filtering with a 5×5 pixel kernel for noise reduction and image binarization based on Otsu's method to obtain a binary mask of the molten pool region. Extract the maximum connected region of the binary mask as the molten pool profile, calculate the width of the minimum bounding rectangle of the molten pool profile as the molten pool width, calculate the height of the minimum bounding rectangle of the molten pool profile as the molten pool length, and calculate the difference between the coordinates of the centroid of the profile perpendicular to the predetermined laser travel direction and the coordinates of the center line of the travel direction as the lateral deviation of the molten pool centroid. The values ​​of width, length, and lateral deviation are arranged in a predetermined order to form a 3×1 column vector, which serves as the feature vector of the three-dimensional molten pool morphology.

3. The method according to claim 1, characterized in that, The control law incorporates a switching function. This function switches between a non-singular fast terminal sliding mode control mode and a continuous high-order polynomial control mode based on whether the norm of the three-dimensional melt pool morphology error vector is greater than a preset switching threshold. It also outputs discrete state signals based on the current control mode, including: Calculate the L2 norm of the three-dimensional molten pool morphology error vector and set a fixed switching threshold; When the L2 norm is not less than the switching threshold, the control law is switched to non-singular fast terminal sliding mode control mode, and a discrete state signal representing the strong response mode is output. When the L2 norm is less than the switching threshold, the control law is switched to a continuous high-order polynomial control mode, and a discrete state signal representing the fine-tuning smoothing mode is output.

4. The method according to claim 1, characterized in that, The parameters of the instruction filter are scheduled by the discrete state signal, including: When the discrete-state signal is in a strong response mode, the boundary layer thickness parameter matrix of the saturation function in the command filter is set as a diagonal matrix, with the diagonal elements being the product of k1 and the channel scaling factor. Simultaneously, the gain coefficient matrix is... , Set as a diagonal matrix, with diagonal elements being the product of k2 and the channel scaling factor; When the discrete state signal is in fine-tuned smoothing mode, the diagonal elements of the boundary layer thickness parameter matrix are set to the product of k3 and the channel scaling factor, and the diagonal elements of the two gain coefficient matrices are set to the product of k4 and the channel scaling factor.

5. The method according to claim 3, characterized in that, The mathematical model used to predict future filtering errors and the prediction time domain length are selected or adjusted by the discrete state signal, including: When the discrete state signal is in a strong response mode, a lightweight recurrent neural network based on a gated recurrent unit is selected to predict the future filtering error, and the prediction time domain length is set to the next 4 control cycles. When the discrete state signal is in a fine-tuned smoothing mode, a second-order linear autoregressive model is selected to predict future filtering errors, and the prediction time domain length is set to the next two control cycles.

6. The method according to claim 1, characterized in that, The process of obtaining three-dimensional control quantities based on the filtered command vector output after processing by the command filter and the compensation signal vector generated by the error compensation system includes: In each control cycle, the filtered virtual control command vector output by the command filter is added to the compensation signal vector output by the error compensation system to obtain the three-dimensional control quantity of the cycle. The three-dimensional control quantity includes laser power, scanning speed and lateral path offset.

7. A system for controlling the flow and solidification of a molten pool on the surface of a nickel-based superalloy, characterized in that, Includes the following modules: The module is used to acquire image information of the molten pool in real time and construct a three-dimensional molten pool morphology feature vector; the difference between the real-time extracted three-dimensional molten pool morphology feature vector and a preset target morphology feature vector is calculated to obtain a three-dimensional molten pool morphology error vector. The output module is used to construct a composite nonlinear control law. Through a mapping relationship including a dimension transformation matrix, it maps and converts the three-dimensional molten pool morphology error vector into a three-dimensional virtual control command vector. The control law has a built-in switching function. The function switches between a non-singular fast terminal sliding mode control mode and a continuous high-order polynomial control mode based on whether the norm of the three-dimensional molten pool morphology error vector is greater than a preset switching threshold. Based on the current control mode, it outputs a discrete state signal. A generation module is used to filter the three-dimensional virtual control command vector using a command filter, the parameters of which are scheduled by the discrete state signal; an error compensation system coupled with the control law mode is established, which generates a compensation signal vector based on the error generated in the current filtering process and the prediction of future filtering errors, wherein the mathematical model used to predict future filtering errors and the prediction time domain length are selected or adjusted by the discrete state signal; and a three-dimensional control quantity is obtained based on the filtered command vector output after processing by the command filter and the compensation signal vector generated by the error compensation system.

8. The system according to claim 7, characterized in that, The real-time acquisition of image information of the molten pool and the construction of a three-dimensional molten pool morphology feature vector include: A camera is used to acquire grayscale images of the molten pool with a resolution of 1280×1024 pixels at a rate of no less than 150 frames per second. The acquired grayscale images are then subjected to Gaussian filtering with a 5×5 pixel kernel for noise reduction and image binarization based on Otsu's method to obtain a binary mask of the molten pool region. Extract the maximum connected region of the binary mask as the molten pool profile, calculate the width of the minimum bounding rectangle of the molten pool profile as the molten pool width, calculate the height of the minimum bounding rectangle of the molten pool profile as the molten pool length, and calculate the difference between the coordinates of the centroid of the profile perpendicular to the predetermined laser travel direction and the coordinates of the center line of the travel direction as the lateral deviation of the molten pool centroid. The values ​​of width, length, and lateral deviation are arranged in a predetermined order to form a 3×1 column vector, which serves as the feature vector of the three-dimensional molten pool morphology.

9. The system according to claim 7, characterized in that, The control law incorporates a switching function. This function switches between a non-singular fast terminal sliding mode control mode and a continuous high-order polynomial control mode based on whether the norm of the three-dimensional melt pool morphology error vector is greater than a preset switching threshold. It also outputs discrete state signals based on the current control mode, including: Calculate the L2 norm of the three-dimensional molten pool morphology error vector and set a fixed switching threshold; When the L2 norm is not less than the switching threshold, the control law is switched to non-singular fast terminal sliding mode control mode, and a discrete state signal representing the strong response mode is output. When the L2 norm is less than the switching threshold, the control law is switched to a continuous high-order polynomial control mode, and a discrete state signal representing the fine-tuning smoothing mode is output.

10. The system according to claim 7, characterized in that, The parameters of the instruction filter are scheduled by the discrete state signal, including: When the discrete-state signal is in a strong response mode, the boundary layer thickness parameter matrix of the saturation function in the command filter is set as a diagonal matrix, with the diagonal elements being the product of k1 and the channel scaling factor. Simultaneously, the gain coefficient matrix is... , Set as a diagonal matrix, with diagonal elements being the product of k2 and the channel scaling factor; When the discrete state signal is in fine-tuned smoothing mode, the diagonal elements of the boundary layer thickness parameter matrix are set to the product of k3 and the channel scaling factor, and the diagonal elements of the two gain coefficient matrices are set to the product of k4 and the channel scaling factor.

Images