3D printing laser power real-time closed-loop control method and system

Abstract

The present application relates to the technical field of laser heat treatment, in particular to a 3D printing laser power real-time closed-loop control method and system, which separates processing laser and thermal radiation signals by using coaxial dichroic reflection elements, extracts multi-band characteristic signals representing three-dimensional morphology and thermal gradient of the molten pool, and performs dynamic light compensation according to the welding surface geometric characteristics. Through spatial filtering and adaptive threshold operator, combined with cross-sectional profile boundary constraint, shadow and highlight noise formed by powder splashing are filtered out, and normalized molten pool characteristic image is obtained. The liquid-solid phase boundary is located, and the instantaneous movement vector of the laser beam and the spatial normal inclination angle are coupled, the characteristic size of the molten pool orthogonal to the scanning track is dynamically calculated. By comparing the target size and introducing the feedforward gain coefficient based on the edge distance of the forming point, nonlinear thermal field compensation and heat treatment regulation are carried out, and power correction instructions are generated. The accumulation of molten droplets and edge spheroidization caused by overheating in the thin-walled area are effectively suppressed, and the geometric distortion induced by welding thermal stress is reduced.

Description

Technical Field

[0001] This invention relates to the field of laser heat treatment technology, specifically to a real-time closed-loop control method and system for 3D printing laser power. Background Technology

[0002] Metal 3D printing, at its microscopic physical essence, involves the layer-by-layer micro-area welding of metal powder, accompanied by in-situ heat treatment processes generated by reciprocating laser scanning. In existing systems encompassing additive manufacturing and intelligent welding, especially laser thermal processing technologies applied to additive manufacturing, welding, and intelligent heat treatment production lines, sensing systems often struggle to eliminate the instantaneous high-brightness noise and shadow interference generated by metal powder spatter, leading to severe distortion in the acquired molten pool characteristic data. Furthermore, traditional detection methods neglect the influence of workpiece surface tilt angle and torsion on imaging geometric distortion, failing to accurately reconstruct the true physical dimensions of the molten pool. Moreover, existing closed-loop controls are mostly lagging feedback adjustments, lacking predictive compensation for heat conduction obstruction induced by geometric positions during continuous heat treatment or unsteady-state heat accumulation processes, easily leading to edge overheating, metal droplet accumulation, and uneven microstructure.

[0003] How to solve the technical problems of severe interference from spatter noise in molten pool sensing, imaging distortion of complex geometric surfaces, and instability caused by edge heat accumulation during the processing (such as micro-area welding).

[0004] To address this, a real-time closed-loop control method and system for 3D printing laser power is proposed. Summary of the Invention

[0005] The purpose of this invention is to provide a real-time closed-loop control method and system for 3D printing laser power. By integrating multi-band sensing dynamic compensation and geometric boundary feedforward control, it aims to achieve sub-pixel-level accurate identification of the molten pool size and real-time power correction under complex thermal field environments.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The radiation signal generated by the processing laser and the melting of metal powder is separated by a dichroic reflective element embedded in the laser coaxial optical path. The band feature signal containing the three-dimensional morphology of the molten pool and the thermal gradient distribution is extracted from the radiation signal, and the sensor is dynamically compensated for photosensitivity based on the geometric features of the metal component surface. Spatial filtering and adaptive threshold operators are used to process the feature signal of the band. The geometric boundary constraints of the profile of the section to be processed are referenced to filter out the shadows formed by metal powder splashes and high brightness noise interference, so as to obtain a normalized molten pool feature image. The outermost liquid-solid phase boundary of the molten metal in the molten pool feature image is located. Combining the direction of the instantaneous movement vector and the spatial normal tilt angle of the component surface torsion, the instantaneous molten pool feature size orthogonal to the current forming direction is dynamically calculated. The energy deviation is calculated by comparing the instantaneous molten pool characteristic size with the target size, and a feedforward gain coefficient is introduced based on the geometric distance from the current forming point to the edge of the component to perform thermal field compensation, thereby generating a power correction command.

[0007] Preferably, extracting the band feature signal includes: The radiation signal is divided into a first sub-beam and a second sub-beam by a beam-splitting component disposed on the reflected light path of the dichroic reflective element; the first sub-beam is guided into a first narrowband filter to obtain a feature image reflecting the edge contour of the molten pool, while the second sub-beam is guided into a second narrowband filter to obtain an image reflecting the thermal gradient characteristics inside the molten pool; a synchronous trigger signal is used to drive a dual sensing unit to synchronously acquire the first sub-beam and the second sub-beam. The dynamic photosensitivity compensation of the sensor based on the geometric features of the metal component surface includes: acquiring the spatial distortion parameters of the current formed surface; and dynamically adjusting the integration time of the sensing unit based on the reflectivity change caused by the spatial distortion parameters, so that the acquired band feature signal is within a preset linear response range.

[0008] Preferably, the step of processing the band feature signal using spatial filtering and an adaptive threshold operator to obtain a normalized molten pool feature image includes: Spatial smoothing filtering is performed on the characteristic signals of each band, and weighted fusion is performed according to the signal-to-noise ratio of each band to obtain a single-channel feature image; the local energy gradient direction in the single-channel feature image is identified, and the complete molten pool core region is extracted by comparing the statistical differences in gradient distribution between the molten pool core region and the metal powder spatter; the pixel grayscale values ​​of the molten pool core region are normalized to generate the normalized molten pool feature image; The geometric boundary constraints of the reference cross-section profile to be processed, filtering out shadows and high-brightness noise interference formed by metal powder splashes, include: acquiring the digital model profile of the cross-section to be processed and mapping it to the imaging coordinate system of the current sensor to construct a dynamic geometric shielding mask; using the dynamic geometric shielding mask to filter out splash shadows and high-brightness outlier noise located outside the geometric boundary and parallel to the component profile.

[0009] Preferably, locating the outermost liquid-solid boundary of the molten metal in the feature image of the molten pool includes: Extract sub-pixel level coordinate points of the edge regions in the normalized molten pool feature image; use a curve fitting algorithm to fit the sub-pixel level coordinate points to construct a closed contour curve representing the outermost liquid-solid phase line of the molten pool. The dynamic calculation of the instantaneous molten pool feature size orthogonal to the current forming direction includes: obtaining the instantaneous movement vector of the laser beam at the current forming point, establishing a spatial follower coordinate system that dynamically rotates and compensates with the laser beam movement; identifying the normal tilt angle between the spatial normal of the component surface at the current forming point and the laser incident direction, and using a cosine compensation algorithm to correct the geometric distortion of the closed contour curve in the longitudinal direction; and calculating, in the spatial follower coordinate system, the maximum chord length in the direction perpendicular to the instantaneous movement vector in the closed contour curve after geometric distortion correction, as the instantaneous molten pool feature size.

[0010] Preferably, the step of introducing a feedforward gain coefficient for thermal field compensation based on the geometric distance from the current forming point to the edge of the component includes: The system identifies thin-walled, high-curvature regions within a component that are subject to limited heat dissipation, and calculates the minimum geometric distance from the current forming point to the edge of the component. It then retrieves a preset nonlinear compensation table to obtain the feedforward gain coefficient corresponding to the minimum geometric distance. The feedforward gain coefficient decreases nonlinearly as the current forming point enters the thin-walled, high-curvature region, with the deceleration rate increasing as the distance decreases, thus suppressing overheating caused by the accumulation of molten metal droplets at the edge. The generation of the power correction command includes: calculating the deviation between the instantaneous molten pool feature size and the preset target size, and calculating the preliminary feedback correction amount using a proportional-integral-derivative control algorithm; weighting the preliminary feedback correction amount with the feedforward gain coefficient to output the final power correction command; and adjusting the output power of the laser through the control interface to make the instantaneous molten pool feature size converge towards the preset target size.

[0011] Preferably, the process of generating the power correction command further includes: The physical properties of the metal powder in the current processing layer are obtained, including the average particle size of the metal powder, the loose packing density of the powder bed, and the surface tension coefficient of the metal powder in the molten state. A molten pool evolution model based on multiphase flow dynamics is established to identify the coupling influence factors of capillary force and metal vapor backflow pressure on the instantaneous molten pool width. Based on the local thermal conductivity distribution of the metal powder bed in the current processing area and combined with the latent heat demand of phase change during the transformation of the metal powder from solid to liquid phase, the feedforward gain coefficient is nonlinearly corrected to compensate for the local heat accumulation fluctuation caused by the uneven particle size of the powder. The spheroidization tendency and spattering loss of metal droplets at the thin-wall edge are suppressed by adjusting the laser power output threshold.

[0012] A real-time closed-loop control system for 3D printing laser power includes: The melting coaxial sensing module separates the radiation signals generated by the processing laser and the melting of metal powder using a dichroic reflective element embedded in the laser coaxial optical path. It extracts band feature signals containing the three-dimensional morphology of the molten pool and the thermal gradient distribution from the radiation signals, and performs dynamic photosensitivity compensation on the sensor based on the geometric features of the metal component surface. The spatter noise reduction module processes the band feature signals using spatial filtering and an adaptive threshold operator. Referring to the geometric boundary constraints of the profile to be processed, it filters out shadows and high-brightness noise interference formed by metal powder spatter, obtaining a normalized molten pool feature image. The molten pool solid-liquid analysis module locates the outermost liquid-solid phase boundary of the molten metal in the molten pool feature image. Combining the direction of the instantaneous movement vector and the spatial normal tilt angle of the component surface's conformal torsion, it dynamically calculates the instantaneous molten pool feature size orthogonal to the current forming direction. The forming thermal field control module compares the instantaneous molten pool feature size with the target size to calculate the energy deviation. Based on the geometric distance from the current forming point to the component edge, it introduces a feedforward gain coefficient for thermal field compensation, generating a power correction command.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By using dichroic beam splitting and multi-band synchronous acquisition technology, combined with dynamic geometric shielding mask, the interference of splash noise and shadows on imaging is effectively eliminated, ensuring sub-pixel-level positioning accuracy of the molten pool liquid-solid phase interface, and providing a real and reliable closed-loop feedback data basis for power regulation.

[0014] 2. An innovative cosine compensation algorithm for spatial normal tilt angle and moving vector is introduced to dynamically correct the imaging geometric distortion caused by complex torsional surfaces, so that the calculated instantaneous molten pool size can truly reflect the forming quality, and solve the measurement deviation problem of traditional visual inspection on non-flat surfaces.

[0015] 3. For thin-walled and high-curvature edges with limited heat dissipation, a feedforward gain coefficient based on geometric distance is introduced to suppress the risk of overheating caused by heat accumulation in advance. Combined with the secondary correction of multiphase flow dynamics parameters, the stability of the molten pool is significantly improved, and forming defects such as edge spheroidization and missing spatter are eliminated. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the steps of a real-time closed-loop control method for 3D printing laser power according to the present invention. Figure 2 This is a schematic diagram of the process for denoising and normalizing the feature image of the molten pool in this invention; Figure 3 This is a system structure diagram of a real-time closed-loop control method for 3D printing laser power according to the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] This invention provides a real-time closed-loop control method and system for 3D printing laser power, referring to... Figure 1 Step flowchart and Figure 3 The system architecture diagram and technical solution are as follows: The radiation signal generated by the processing laser and the melting of metal powder is separated by a dichroic reflective element embedded in the laser coaxial optical path. The band feature signal containing the three-dimensional morphology of the molten pool and the thermal gradient distribution is extracted from the radiation signal, and the sensor is dynamically compensated for photosensitivity based on the geometric features of the metal component surface. Spatial filtering and adaptive threshold operators are used to process the feature signal of the band. The geometric boundary constraints of the profile of the section to be processed are referenced to filter out the shadows formed by metal powder splashes and high brightness noise interference, so as to obtain a normalized molten pool feature image. The outermost liquid-solid phase boundary of the molten metal in the molten pool feature image is located. Combining the direction of the instantaneous movement vector and the spatial normal tilt angle of the component surface torsion, the instantaneous molten pool feature size orthogonal to the current forming direction is dynamically calculated. The energy deviation is calculated by comparing the instantaneous molten pool characteristic size with the target size, and a feedforward gain coefficient is introduced based on the geometric distance from the current forming point to the edge of the component to perform thermal field compensation, thereby generating a power correction command.

[0019] Example 1: This embodiment is set in the laser additive manufacturing scenario of turbine metal blades in the aerospace field, and specifically implements a method for real-time monitoring and power closed-loop control of the melting process using optical components embedded in the laser coaxial optical path.

[0020] The layer-by-layer stacking process in metal 3D printing is essentially a high-frequency, continuous micro-area welding process. In intelligent welding systems, the morphological stability of the molten pool and the evolution of the liquid-solid phase interface directly determine the mechanical properties and density of the weld joint and the deposited layer. The laser power closed-loop control method proposed in this invention can monitor and provide feedback on the melt flow and transient heat transfer behavior during the micro-area welding process in real time. By dynamically adjusting the heat input, it effectively avoids metallurgical defects such as incomplete fusion, micropores, and hot cracks that are prone to occur in traditional multi-layer, multi-pass welding, thereby ensuring the internal microstructure uniformity and quality consistency of complex metal components during the layer-by-layer micro-welding process.

[0021] Furthermore, a dichroic reflective element embedded in the laser coaxial optical path is used to separate the processing laser from the radiation signal generated by the melting of metal powder. Specifically, during the printing of alloy blades resistant to high temperature gradients, the system places a dichroic reflective element with an installation angle of 45 degrees between the laser output head and the scanning galvanometer, allowing the processing laser beam with a wavelength of 1064nm to pass directly through the element and be focused on the surface of the metal powder bed. At the same time, the reflective coating of the element reflects the broadband radiation signal with a wavelength range of 400nm to 2500nm generated by the molten pool during the melting process to the monitoring branch at a 90-degree angle. This achieves decoupling and separation of the high-energy processing beam and the low-energy sensing signal in the physical path, ensuring that the sensor obtains real-time feedback data with a high signal-to-noise ratio without direct interference from the laser.

[0022] Furthermore, band feature signals containing the three-dimensional morphology and thermal gradient distribution of the molten pool are extracted from the radiation signal. Specifically, the reflected radiation signal enters the feature extraction module through a set of achromatic collimating lenses. The system uses an optical beam splitter with a splitting ratio of 50:50 to guide the light signal into two parallel detection channels. The first channel acquires a 512×512 pixel resolution geometric image of the molten pool through an industrial sensor with a sampling rate of 5000Hz. The second channel acquires the infrared radiation intensity matrix of the molten pool surface through a high-speed sensor with a thermal sensitivity of 0.1 degrees Celsius. These raw data streams are transmitted in real time to the memory of the local industrial controller via the PCIe 3.0 bus, providing basic data support for subsequent calculations of the aspect ratio and temperature field distribution of the molten pool.

[0023] Furthermore, by utilizing a beam splitter component disposed on the reflective optical path of the dichroic reflective element, the radiation signal is divided into a first sub-beam and a second sub-beam. Specifically, the beam splitter component employs a semi-transparent, semi-reflective mirror with multiple dielectric films to evenly distribute the radiation energy entering the monitoring optical path. The first sub-beam propagates along the horizontal optical axis and passes through a first narrowband filter with a center wavelength of 850 nm and a bandwidth of 20 nm to filter out the interference of blue-green plasma light generated by the evaporation of metal powder, thereby obtaining a molten pool edge contour feature image with a contrast of over 95%. The second sub-beam is refracted at 90 degrees and passes through a second narrowband filter with a center wavelength of 1300 nm and a bandwidth of 50 nm to lock the heat flow radiation characteristics in the near-infrared band, thereby obtaining a grayscale image reflecting the internal thermal gradient characteristics of the molten pool.

[0024] Specifically, the dichroic reflective element employs a customized multilayer dielectric film design with the following spectral characteristics: a high-transmittance window at 1064nm wavelength, with a transmittance greater than 98% and a full width at half maximum (FWHM) of approximately 20nm (i.e., maintaining high transmittance within the 1054nm to 1074nm range), ensuring that the processing laser can pass through the element without loss; and high-reflectance regions in the 400nm to 1050nm and 1080nm to 2500nm wavelength ranges, with an average reflectance greater than 85%, to collect thermal radiation and plasma radiation signals generated by the molten pool. By designing a narrow-band transmission window near the 1064nm wavelength, the system effectively avoids interference from high-intensity lasers on the monitoring optical path while retaining key radiation signals characterizing the molten pool state. It is important to note that the reflected laser at 1064nm wavelength (laser reflected from the powder bed surface) is collinear with the incident laser and will return along the original optical path to the laser direction, without entering the monitoring branch, thus avoiding saturation interference to the sensor. The monitoring branch mainly collects high-temperature metal radiation from the molten pool (primarily concentrated in the near-infrared and visible light bands) and plasma radiation (ultraviolet and blue-green light bands), all of which are within the high reflectivity zone of the dichroic mirror.

[0025] Furthermore, the system utilizes a synchronous trigger signal to drive dual sensing units to synchronously acquire the first and second sub-beams. Specifically, the system uses a timing controller based on an FPGA architecture to generate a synchronous TTL pulse signal with a frequency of 10kHz. This signal is simultaneously connected to the trigger interfaces of the two CMOS sensing units via shielded cables, ensuring that the electronic shutters of the sensors are opened synchronously within the time window. The acquired first sub-beam image contains the melt width and flow vector characteristics of the molten pool, while the second sub-beam image corresponds to the instantaneous thermal equilibrium distribution at the same location. The digital image sequences output by the two sensing units are stored in the system's dual-port random access memory. By calculating the spatiotemporal correlation between pixels, a time-frequency synchronized closed-loop data input is provided for the joint analysis of the molten pool state.

[0026] Furthermore, dynamic photosensitivity compensation is performed on the sensor based on the geometric features of the metal component surface. Specifically, the system deploys a deep learning model based on a CNN architecture to perform photosensitivity strategy prediction. This CNN model consists of an input layer, three convolutional layers with a kernel size of 3×3, two pooling layers of 2×2, and a fully connected layer containing 128 neurons. The model is input with spatial distortion parameters such as the curvature distribution and normal vector deflection of the blade surface in the current printing layer. The convolutional layers extract the local spatial correlation of geometric features and use the ReLU function for nonlinear activation, finally outputting a compensation factor ranging from 0 to 1. When training this model, the system uses the Adam optimizer and sets a learning rate of 0.001. The weights are iteratively updated by minimizing the mean square error between the predicted optimal integration time and the labeled true integration time.

[0027] Specifically, the detailed architecture of the CNN model is as follows: The input layer receives a grayscale image with dimensions of 64×64×1. This image is generated by rasterizing and downsampling the 3D point cloud data of a 5mm×5mm area surrounding the current printing position, where the pixel value represents the normalized angle between the local surface normal vector and the horizontal plane. The first convolutional layer contains 32 3×3 convolutional kernels with a stride of 1, using a same padding strategy to maintain the feature map size, and outputs a 64×64 feature map with 32 channels. The convolutional layer is followed by a ReLU activation function and a batch normalization layer. The first pooling layer uses 2×2 max pooling with a stride of 2, reducing the feature map size to 32×32. The second convolutional layer contains 64 3×3 convolutional kernels with a stride of 1, using same padding, and outputs a 32×32 feature map with 64 channels. It is also followed by a ReLU and batch normalization layer. The second pooling layer uses 2×2 max pooling with a stride of 2, reducing the feature map size to 16×16. The third convolutional layer contains 128 3×3 convolutional kernels with a stride of 1 and same padding, outputting a 16×16 feature map with 128 channels, followed by ReLU and batch normalization layers. Global average pooling then compresses the 16×16×128 feature map into a 128-dimensional feature vector. This vector is input to a fully connected layer containing 128 neurons. The fully connected layer uses the Sigmoid activation function and outputs a single numerical value as a compensation factor, ranging from 0 to 1.

[0028] The model was trained using 2000 sets of labeled data, sourced from molten pool images and their corresponding optimal integration time labels acquired under different surface tilt angles (0° to 60°) and height deviations (-2mm to +2mm). During training, mean squared error (MSE) was used as the loss function, comprising both the model-predicted compensation factor and an ideal compensation factor calculated based on the actual image saturation (by statistically analyzing the proportion of pixels with values ​​greater than 250 in the image histogram; when this proportion exceeds 5%, it is considered overexposed, and the compensation factor should be decreased; when the proportion is below 0.1%, it is considered underexposed, and the compensation factor should be increased). Training employed the Adam optimizer with a learning rate of 0.001, a batch size of 32, and 100 training epochs. Training stopped when the MSE loss on the validation set no longer decreased for 10 consecutive epochs.

[0029] Furthermore, the system acquires the spatial distortion parameters of the current forming surface and dynamically adjusts the integration time of the sensing unit. Specifically, the system extracts the surface tilt angle and height deviation data of the current printing position in real time from the slice layer description file, and calculates the light intensity fluctuation component by combining it with the local reflectivity of the metal surface. Based on the reflectivity change caused by the spatial distortion parameters, the FPGA controller adjusts the integration time of the sensing unit in real time to ensure that the band feature signal acquired at the blade edge or steep area is always within the linear response range of 20% to 80% of the sensor's dynamic range, avoiding image overexposure or loss of detail, thereby providing a stable and reliable image feature input for precise closed-loop control of laser power.

[0030] As an implementable approach, the multidimensional thermal gradient vector distribution in the molten pool feature image is calculated in real time to identify the micro-region topological features in the radiation signal that characterize the fluctuation of the metal droplet's absorption rate to the laser; the real-time offset of the laser-matter energy coupling efficiency is calculated based on the micro-region topological features, and the offset is used to fine-tune the power correction command with feedback polarity.

[0031] During processing, considering the dynamic balance between the surface tension and steam back pressure of the metal powder at the moment of melting, non-stationary fluctuations will occur on the surface of the molten pool. The system analyzes the intensity contrast changes of the first and second sub-beams to map the micro-roughness distribution of the molten pool surface. When the micro-roughness exceeds a preset threshold, causing a decrease in the effective laser absorption rate, the control algorithm adds a positive pulse component proportionally to the original power correction command to compensate for the effective energy input loss due to the reduced absorption rate. This solves the inherent bias of the default constant laser absorption rate in traditional control. By sensing the physical morphology of the molten pool surface and correcting the absorption rate fluctuations in real time, the system achieves energy input compensation and significantly reduces fusion defects caused by surface disturbances.

[0032] By decoupling the acquisition of multi-band radiation signals through a beam splitter, and using narrowband filtering technology to extract the edge contour and internal thermal gradient features of the molten pool, the problem of missing single-band sensing information is solved. Combined with dynamic integral time compensation based on spatial distortion parameters, this ensures that the sensing unit remains within the linear response range when processing complex reflective surfaces, preventing molten pool feature distortion caused by signal saturation or underexposure from the source, and improving the robustness of photoelectric sensing.

[0033] Furthermore, referring to Figure 2 A schematic diagram of the denoising and normalization process for molten pool feature images is provided. The step of processing the band feature signals using spatial filtering and an adaptive threshold operator to obtain a normalized molten pool feature image includes: performing spatial smoothing filtering on each band feature signal and weighted fusion based on the signal-to-noise ratio of each band to obtain a single-channel feature image; identifying the local energy gradient direction in the single-channel feature image; extracting the complete connected molten pool core region by comparing the statistical differences in gradient distribution between the molten pool core region and the metal powder spatter; and normalizing the pixel grayscale values ​​of the molten pool core region to generate the normalized molten pool feature image.

[0034] Specifically, the adaptive threshold operator employs an adaptive segmentation method based on local gray-level statistical characteristics. The system divides the single-channel feature image into several 16×16 pixel local windows, with the overlap rate between windows set to 50% to ensure boundary continuity. For each local window, the mean μ and standard deviation σ of its internal pixel gray-level values ​​are calculated, and the local threshold T of the window is calculated according to the formula T=μ+k×σ, where k is an adaptive adjustment coefficient, dynamically set according to the ratio of the current window's gray-level variance to the overall image, with a value range of 0.5 to 1.5. When the gray-level variance within the window is large (indicating the presence of melt pool edges or spatter), the value of k is automatically increased to improve threshold sensitivity; when the variance is small (indicating a uniform background area), the value of k is decreased to avoid noise misjudgment. The gray-level value of each pixel is compared with the local threshold T of its window; pixels greater than the threshold are marked as foreground (melt pool region), and pixels less than the threshold are marked as background, thereby achieving adaptive extraction of melt pool features.

[0035] Specifically, the graphics processing unit of the industrial controller first retrieves a sequence of 512×512 pixel radiation images synchronously acquired by the dual sensing units from the dual-port random access memory.

[0036] For the received multi-band image data, the system uses a 3×3 spatial convolution kernel to perform smoothing filtering on the feature signal of each frame. The weight parameters inside the convolution kernel are set according to the discrete point spread function, thereby eliminating the granular noise in the image caused by the dark current of the sensor.

[0037] Furthermore, the system calculates the specific signal-to-noise ratio (SNR) of each band in real time based on the ratio of the global variance to the mean of the feature image of each band. Using a fusion weight generator based on Softmax logic, the SNR values ​​of each band are converted into weight coefficients with a sum of 1 and assigned to the corresponding channels. Specifically, the weighted score is obtained by taking the exponent with the natural constant as the base of each SNR value and then performing normalization. Subsequently, the first sub-beam image and the second sub-beam image are linearly superimposed according to this weight to obtain a single-channel feature image with a dynamic range of more than 60dB.

[0038] Specifically, the steps for comparing the statistical differences in gradient distribution between the molten pool core area and the metal powder spatter are as follows: The system first uses the Sobel operator to perform full-field gradient calculation on the single-channel feature image to obtain the gradient magnitude G and gradient direction θ of each pixel. For any connected region in the image, the distribution histogram of the gradient direction of all pixels within it in the range of 0 to 360 degrees is calculated, the direction is divided into 36 intervals (each interval is 10 degrees), and the proportion of pixels in each interval is counted.

[0039] The molten pool core region, due to its circular or elliptical shape as a pool of liquid metal, exhibits significant radial symmetry in its gradient direction. That is, the gradient direction is uniformly distributed radially outward from the center of the molten pool, and the distribution of the direction histogram is relatively uniform across all intervals, with a small standard deviation σ_pool. In contrast, metal powder spatter, due to its irregular shape and discrete distribution formed by random splashing, exhibits a highly random gradient direction, with the direction histogram showing sharp peaks in certain intervals and a larger standard deviation σ_spatter.

[0040] The system calculates the standard deviation σ of the gradient direction histogram for each candidate connected region and compares it with the preset discrimination threshold σ_threshold=0.25: when σ<σ0, it is determined to be the core region of the molten pool; when σ>σ_threshold, it is determined to be the spatter noise region and is filtered out. Furthermore, the system incorporates the area constraint of connected regions, classifying tiny connected regions with an area less than 50 pixels as noise. Through the differential discrimination based on the aforementioned gradient distribution statistical characteristics, the system can accurately extract the complete core region of the molten pool from the connected regions. The discrimination threshold of 0.25 was determined by performing discrete statistical calibration on 5000 historical samples of molten pool and spatter. The system uses a bimodal search algorithm to fit the probability density distribution curves of the molten pool core region (mean standard deviation 0.12) and spatter noise (mean standard deviation 0.41). Based on the statistical intersection point of the two feature distributions at 0.21, an additional 15% safety weight increment is added to compensate for the fluctuations in reflectivity of different metal materials and environmental noise. Thus, 0.25, which balances recognition accuracy and system robustness, is ultimately selected as the standard deviation discrimination threshold.

[0041] The system constructs a lightweight CNN network containing four convolutional layers and two skip connection layers to perform the extraction of the molten pool core region. The input layer of the network receives the single-channel feature image mentioned above. The first convolutional layer uses 16 5×5 convolutional kernels to extract large-scale energy distribution features, and the second convolutional layer uses 32 3×3 convolutional kernels to finely identify the local energy gradient direction. By analyzing the statistical difference between the radially symmetrical gradient distribution exhibited by the molten pool core region and the random high-frequency gradient exhibited by the metal powder spatter, the network outputs a probability distribution map with the same size as the original image. The value of each pixel in the probability map represents its confidence score of belonging to the molten pool core region.

[0042] Furthermore, the system binarizes the probability distribution map by setting a probability threshold of 0.85 and uses a morphological dilation operator to close the broken edges, thereby accurately locating the complete molten pool core region in a single-channel image. The probability threshold of 0.85 is determined by balancing the precision-recall curve of the CNN model on the validation set. During the training phase, the system uses the Softmax activation function to output the confidence distribution of each pixel. By statistically analyzing the inference results of 10,000 validation set samples, it was found that when the threshold is set to 0.85, the false positive rate of the model misclassifying spatter noise as the molten pool core region can be stably controlled below 0.5%, while ensuring that the edge feature recall rate of the molten pool core region reaches above 98%. This value aims to prioritize the purity of the input source of the closed-loop control system through high-confidence logical constraints, avoid laser power misadjustment caused by high-dynamic spatter interference, and reserve approximately 10% robust redundancy space for possible dust contamination or optical attenuation in the optical path system.

[0043] The system performs Min-Max normalization on the extracted core region of the molten pool. By retrieving the maximum and minimum gray values ​​within the current region of interest, the system subtracts the minimum value from the original gray value of each pixel and divides it by the gray range value, thereby uniformly mapping the image brightness to a dimensionless numerical space of 0 to 1. This generates a contrast-enhanced normalized molten pool feature image, which is then output to the subsequent geometric denoising module.

[0044] Furthermore, the geometric boundary constraint of the reference cross-section profile to be processed, filtering out shadows and high-brightness noise interference formed by metal powder splashes, includes: acquiring the digital model profile of the cross-section to be processed and mapping it to the imaging coordinate system of the current sensor to construct a dynamic geometric shielding mask; using the dynamic geometric shielding mask to filter out splash shadows and high-brightness outlier noise located outside the geometric boundary and parallel to the component profile.

[0045] Specifically, the system extracts the STL digital model outline of the blade section to be processed in the current layer from the layer thickness slice database of the host computer main controller in real time. The outline data is stored in a vector format containing a set of three-dimensional coordinate points.

[0046] The system uses a pre-calibrated 4×4 affine transformation matrix to align the spatial coordinate system of the laser scanning galvanometer with the imaging coordinate system of the CMOS sensor, and projects the vector contour data onto the current imaging plane, thereby fitting the actual processing boundary of the blade cross section onto the sensor's image plane.

[0047] Furthermore, the system constructs a dynamic geometric mask based on the geometric boundary obtained by projection. The mask is a binary matrix that is completely consistent with the imaging resolution, wherein the pixel value inside the physical area of ​​blade processing is set to 1, while the pixel value of the background area outside the processing boundary is set to 0.

[0048] The system performs a pixel-by-pixel logical AND operation on the normalized molten pool feature image and the dynamic geometric mask. Since high-temperature metal spatter often flies out of the blade outline when printing the blade edge, generating a large amount of outlier brightness interference, the system can automatically filter out spatter shadows located outside the geometric boundary and parallel to the component outline, as well as all high-brightness outlier noise, through the physical blocking effect of the mask, ensuring that the molten pool features retained in the end come only from the actual printing area.

[0049] The high-purity molten pool image after geometric constraint denoising is used as the input feature of the closed-loop controller. The system periodically issues power adjustment commands based on the pixel centroid position and area fluctuation of the retained area, thereby ensuring the consistency of the melting quality of the turbine blades during the laser additive manufacturing process.

[0050] Furthermore, the system locates the outermost liquid-solid phase boundary of the molten metal in the molten pool feature image. Combining the direction of the instantaneous movement vector and the spatial normal tilt angle of the component surface torsion, it dynamically calculates the instantaneous molten pool feature size orthogonal to the current forming direction. Specifically, the central processing unit of the industrial controller first receives the normalized molten pool feature image from the previous module and calls the sub-pixel edge extraction algorithm deployed in the graphics processor. By performing bicubic interpolation on the transition region with drastic changes in brightness gradient in the image, the original 512×512 resolution pixel array is expanded to an equivalent 5120×5120 in the local edge region, achieving sub-pixel level edge extraction and equivalent spatial positioning. With a bit accuracy better than 0.1 pixels, the system determines the set of sub-pixel level coordinate points at the 0.5 boundary line for the grayscale value of the molten pool edge. Subsequently, the system obtains the laser head movement vector data fed back in real time by the CNC system through the bus interface. This data contains the three-dimensional components of the current movement direction in the workspace coordinate system. Combined with the surface normal vector preset by the forming model slice data, a follow-up feature coordinate system that is always dynamically aligned with the laser beam movement direction is established using a coordinate rotation operator. Under this coordinate system, the geometric deformation caused by the torsion of the blade surface is offset by projection transformation. Finally, the maximum geometric span of the molten pool in the width direction is calculated and output as the instantaneous molten pool feature size to the power compensation controller.

[0051] Furthermore, sub-pixel level coordinates of edge regions in the normalized molten pool feature image are extracted. Specifically, the system uses a preset 3×3 Sobel operator to perform a full-field gradient scan on the image, identifies candidate pixel regions with gray-level gradient magnitudes greater than 0.6, and then performs surface fitting based on second-order Taylor expansion for each candidate pixel. By finding the extreme points of the local gradient function, the precise position of the edge is located, improving the edge positioning accuracy from 1 pixel unit to 0.1 pixel unit. The output is a list containing 500 to 800 sub-pixel level coordinates, each carrying high-precision two-dimensional floating-point coordinates.

[0052] Furthermore, a curve fitting algorithm is used to fit the sub-pixel level coordinate points to construct a closed contour curve representing the outermost liquid-solid phase line of the molten pool. Specifically, the system constructs a cubic B-spline fitting operator based on the least squares method, using the sub-pixel level coordinate points obtained above as input constraints for control points. By performing iterative solutions to a large-scale linear equation system in the shared memory of the graphics processor, a continuous and closed parameterized curve is constructed. This curve completely outlines the boundary between the liquid metal and the solid matrix in the molten pool during the metal powder smelting process. Its goodness of fit reaches over 98%, and the output is a set of spline curve equations defined by 20 control points.

[0053] Furthermore, the instantaneous movement vector of the laser beam at the current forming point is obtained, and a spatial servo coordinate system that dynamically rotates and compensates with the movement of the laser beam is established. Specifically, the system reads the instruction data stream of the CNC system in real time through a high-speed industrial Ethernet interface to obtain the instantaneous velocity component of the laser scanning galvanometer at the current forming position. For example, when printing a thin-walled blade structure, the speed is set to 800 mm / s. The system uses a 2×2 rotation transformation matrix to map the static pixel coordinate system where the image sensor is located to a dynamic servo coordinate system with the direction of the movement vector as the positive X-axis, thereby eliminating the directional interference caused by the frequent switching of the laser head movement direction to the molten pool morphology discrimination.

[0054] Furthermore, the system identifies the normal tilt angle between the spatial normal of the component surface at the current forming point and the laser incident direction. A cosine compensation algorithm is used to correct the geometric distortion of the closed contour curve in the longitudinal direction. Specifically, the system constructs an angle regression neural network based on the ResNet-18 architecture. This model contains 18 convolutional layers. The input is a multi-band fused image acquired by dual sensing units. After feature extraction by a large 7×7 convolutional kernel and four residual block groups, a scalar value representing the current surface tilt angle is output by a fully connected layer. During the training phase, the model uses the Adam optimizer and is pre-trained using 10,000 sets of simulated molten pool images with known normal tilt angles. The output angle value is multiplied by the system's built-in 1064nm laser incident vector to obtain the normal tilt angle. Based on the cosine value of this angle, the system performs longitudinal stretching compensation on the closed contour curve in the following coordinate system, restoring the elliptic projection distortion caused by the blade's torsional surface to the true physical morphology distribution.

[0055] The cosine compensation algorithm is applicable when the normal tilt angle θ is in the range of 0 to 45 degrees. Specifically, the system calculates the cosine correction factor based on the tilt angle value θ output by the ResNet-18 network. The coordinates of the closed contour curve in the longitudinal direction (i.e., the laser incident direction) are scaled: z_c = z_m / cos(θ), where z_m is the measured longitudinal coordinate and z_c is the corrected true coordinate. When the tilt angle θ is less than 45 degrees, cosine compensation can control the measurement error within 3%, meeting the accuracy requirements for molten pool size control.

[0056] When the system detects a local tilt angle θ exceeding 45 degrees (such as the trailing edge tip of a blade or a region of sharp twist), simple cosine compensation is insufficient to eliminate projection distortion. At this point, the system automatically switches to a binocular stereo vision-based 3D reconstruction mode: utilizing dual-view images from the first and second sensing units, it calculates the 3D point cloud of the molten pool surface using a stereo matching algorithm (such as SGBM semi-global matching), directly extracting the true 3D dimensions of the molten pool from the point cloud data without relying on geometric corrections from the 2D projection. This dual-mode switching mechanism ensures the system's measurement accuracy across the entire angular range.

[0057] Furthermore, in the spatial servo coordinate system, the maximum chord length in the direction perpendicular to the instantaneous movement vector in the closed contour curve after geometric distortion correction is calculated as the instantaneous molten pool characteristic dimension. Specifically, the system performs scanning optimization along the normal direction orthogonal to the current movement vector direction in the corrected closed contour curve equation. By calculating the Euclidean distance between the two intersection points with the same X coordinate and the largest Y-axis distance on the curve, the instantaneous melt width data of the molten pool is extracted. For example, when processing the edge of a blade with a thickness of 2mm, this value fluctuates between 0.5mm and 0.8mm in real time. This characteristic dimension is fed back to the power PID regulator of the control system as a single-precision floating-point number with a frequency of 500Hz, which is used to dynamically adjust the output power of the laser and realize precise closed-loop control of the heat input during additive manufacturing.

[0058] A precise closed contour model of the molten pool's liquid-solid phase line was established through sub-pixel-level edge extraction and curve fitting. By coupling the instantaneous laser movement vector with the spatial normal tilt angle, geometric distortions caused by non-orthogonal laser incidence and component surface torsion were eliminated, enabling dynamic measurement of the instantaneous physical dimensions of the molten pool. This method solves the projection offset problem of traditional visual inspection on non-flat forming surfaces, significantly improving the geometric accuracy of molten pool characterization in three-dimensional space.

[0059] Furthermore, the system compares the instantaneous molten pool feature size with the target size to calculate the energy deviation, and introduces a feedforward gain coefficient based on the geometric distance from the current forming point to the component edge for thermal field compensation, generating a power correction command. Specifically, the industrial controller extracts a preset target molten width value of 0.65mm from memory as a reference value, performs a difference operation with the instantaneous molten pool feature size with single-precision floating-point format output by the previous module, and obtains the size deviation value under the current thermal input state. At the same time, the system reads the encoder feedback of the laser galvanometer in real time, locks the position index of the current scanning point in the current slice layer in the 512×512 pixel image plane coordinate system, and calls the pre-stored component edge vector path, uses the Euclidean distance operator to calculate the normal geometric span from the current point to the nearest boundary path, thereby generating a feedforward control command to compensate for the heat accumulation effect at the blade edge.

[0060] Furthermore, the system identifies thin-walled, high-curvature regions within the component that are subject to heat dissipation constraints and calculates the minimum geometric distance from the current forming point to the edge of the component. Specifically, the system constructs a geometric feature analysis model based on a GCN (Graph Convolutional Network) architecture. This model includes one input layer, two graph convolutional layers, and one global pooling layer. The input layer receives graph structure data containing the topological relationships of the blade slice contours, where nodes represent discrete sampling points on the contours, and edges represent the connection relationships and curvature features between sampling points. The graph convolutional layers aggregate the features of adjacent nodes to calculate the topological compactness and radius of curvature of each local area. The model uses Softmax normalization logic to assign a heat-constrained weight to each node. Specifically, the weight is calculated by calculating the exponential mapping ratio of the curvature values ​​of each region to obtain a value between 0 and 1. When the weight value exceeds 0.85, it is identified as a thin-walled, high-curvature region at the trailing edge or tip of the blade.

[0061] Furthermore, the system acquires graph structure data containing the topological relationships of the blade slice contours. Specifically, the construction process of the graph structure data first involves the industrial controller's geometry engine extracting the closed contour curve of the blade cross-section from the digital model of the slice layer, and uniformly sampling along the curve at equal arc length intervals of 0.5 mm to generate a node set containing 100 to 500 discrete sampling points. Each node is assigned a four-dimensional feature vector. The specific content of the feature vector includes the normalized pixel horizontal and vertical coordinates, the local curvature calculated using a three-point circle fitting algorithm, and the radial distance from the point to the geometric centroid of the contour. The calculation logic of the local curvature is to obtain the coordinates of the current node and its adjacent nodes and calculate the area of ​​the triangle formed by the three. Then, four times the area of ​​the triangle is divided by the product of the lengths of the three sides to obtain the curvature feature value that reflects the sharpness of the blade edge. All generated node feature data are temporarily stored in the graphics processor memory of the industrial controller.

[0062] Furthermore, the edge connections in the graph structure data follow a rule that combines basic circular connections with spatial proximity connections. Specifically, the system first connects each node to its preceding and following adjacent nodes along the contour path to establish basic edges that reflect the geometric topological continuity. Then, the parallel computing unit of the industrial controller calculates the Euclidean distance between each node and all other nodes. If the physical distance between two points is less than a preset threshold of 2.0 mm, additional spatial proximity edges are established in the video memory. This model in the graph data creates thin-walled features on both sides of a slender branch structure that are close in space but far apart in contour. Each edge carries a two-dimensional feature vector, which includes the Euclidean distance between two connected nodes and the absolute value of the difference in local curvature between the two nodes.

[0063] Furthermore, the system utilizes a GCN architecture comprising two graph convolutional layers and one global pooling layer to predict geometric features of graph structure data. Specifically, in each graph convolutional layer, the node feature update process is achieved by aggregating the feature information of neighboring nodes. The specific logic is to multiply the feature vector of each neighboring node by a pre-set trainable weight matrix in GPU memory, and then normalize the weights based on the reciprocal of the square root of the product of the degree of the current node and its neighboring nodes. After weighted summation, the node features are nonlinearly transformed using the ReLU activation function to extract deep topological relationships. The node features after two layers of graph convolution processing are then fed to the global average pooling layer. The arithmetic mean of the features of all nodes in the graph is calculated to generate a graph-level global feature representation. Finally, this feature representation is input into the Softmax classifier to calculate and output the specific probability value of each sampling point belonging to a heat-constrained thin-walled high-curvature region. This probability data is then passed to the subsequent feedforward gain calculation module to perform thermal field compensation.

[0064] Furthermore, the system identifies thin-walled, high-curvature regions within the component that are subject to heat dissipation constraints and calculates the minimum geometric distance from the current forming point to the edge of the component. Specifically, the system logically compares the probability value of each sampling point output by the GCN architecture with a preset judgment threshold of 0.85. When the probability value of a specific node exceeds the threshold, it is marked as a thin-walled, high-curvature region with limited heat dissipation. Subsequently, the system utilizes the geometric distance calculation unit deployed in the field-programmable gate array to perform real-time simultaneous equations of the scanning coordinates of the current laser beam and the curve equation of the slice contour. By finding the minimum value of the distance functional, the system accurately calculates the minimum geometric distance from the current forming point to the nearest contour edge. This value is output to the back-end gain retrieval module in 12-bit binary encoding format.

[0065] Specifically, after identifying the heat dissipation-limited area, the system uses the geometric distance calculation unit deployed in the FPGA chip of the industrial control motherboard to combine the scanning coordinates of the current laser beam with the B-spline curve equation of the slice contour in real time. By finding the minimum value of the distance functional, the minimum geometric distance from the current forming point to the edge of the component is accurately calculated. This value is output to the gain retrieval module at the back end in 12-bit binary encoding format.

[0066] Furthermore, a preset nonlinear compensation table is retrieved to obtain the feedforward gain coefficient corresponding to the minimum geometric distance. Specifically, the system pre-stores a nonlinear lookup table containing 256 sampling points in non-volatile memory. This table is obtained through multiple thermal simulations and single-channel experiments on the blade edge and reflects the mapping relationship between edge heat dissipation conditions and the required laser power attenuation ratio. When the minimum geometric distance is greater than 5.0 mm, the system outputs a feedforward gain coefficient with a constant of 1.0. When the minimum geometric distance enters the edge transition zone of 0.5 mm to 2.0 mm, the system executes the table lookup instruction and uses a linear interpolation algorithm to calculate the corresponding coefficient value, ensuring that the adjustment of laser energy has physical continuity.

[0067] Specifically, the generation process of the nonlinear compensation table is as follows: First, a three-dimensional transient heat conduction model of the blade edge is established using ANSYS finite element simulation software. In the model, the metal powder bed adopts the equivalent thermal conductivity of porous medium (5W / m·K), and the metal matrix adopts the measured thermal conductivity (20W / m·K). A moving Gaussian heat source (power 350W, spot diameter 0.8mm, scanning speed 800mm / s) is applied to the molten pool region. The steady-state molten pool width is calculated under different edge distances (from 0.1mm to 10mm, a total of 100 distance points) and compared with the target molten width of 0.65mm. When the simulated molten width exceeds the target value, the input power is gradually reduced until the molten width converges to the target value. The power reduction ratio at this time is recorded as the feedforward gain coefficient at that distance point.

[0068] Secondly, ten sets of thin-walled samples with different wall thicknesses (0.5 mm to 3 mm) were prepared on an actual printing platform. Twenty temperature measurement points were evenly arranged along the wall thickness direction for each set of samples. The temperature history of each measurement point was simultaneously acquired using an infrared thermal imager during the scanning process. Peak temperatures at different distances from the edge were extracted and compared with the target temperature (50°C above the melting point) to calculate the power adjustment coefficient required to reach the target temperature. The simulation data and experimental data were weighted and fused (simulation data weight 0.6, experimental data weight 0.4), and a nonlinear lookup table with 256 uniformly sampled points was generated using cubic spline interpolation.

[0069] The following is a sample of the data in this table: When the geometric distance is 0.1 mm, the feedforward gain coefficient is 0.70; When the geometric distance is 0.5 mm, the feedforward gain coefficient is 0.82; When the geometric distance is 1.0 mm, the feedforward gain coefficient is 0.90; At a geometric distance of 2.0 mm, the feedforward gain coefficient is 0.95; When the geometric distance is 5.0 mm or more, the feedforward gain coefficient is 1.00; When the real-time geometric distance input to the system is between two sampling points, the linear interpolation formula is used to calculate: K = K1 + (K2 - K1) × (d - d1) / (d2 - d1), where K is the output feedforward gain coefficient, K1 and K2 are the coefficient values ​​of two adjacent sampling points, d is the input geometric distance, and d1 and d2 are the distance values ​​of two adjacent sampling points. When the input distance is less than 0.1 mm, the system clamps it to 0.1 mm and outputs the corresponding coefficient 0.70; when the input distance is greater than 10 mm, the system clamps it to 10 mm and outputs the coefficient 1.00.

[0070] Specifically, when the current forming point enters the thin-walled high-curvature region, the feedforward gain coefficient decreases nonlinearly with the decrease of the minimum geometric distance, and the deceleration rate increases with the decrease of distance. Specifically, when the minimum geometric distance decreases from 1.0 mm to 0.2 mm, the feedforward gain coefficient called by the system decreases nonlinearly from 0.95 to 0.70. Through this negative gain compensation, the system can automatically reduce the output current of the laser, thereby effectively suppressing the local overheating and geometric collapse phenomenon caused by the obstruction of heat conduction at the ultra-thin edge of the blade, ensuring that the shape accuracy of the component edge reaches more than 98%.

[0071] Furthermore, the generation of power correction instructions includes: calculating the deviation between the instantaneous molten pool characteristic size and the preset target size, and calculating the preliminary feedback correction amount by combining the proportional-integral-derivative control algorithm. Specifically, the system constructs a negative feedback adjustment loop based on the PID control architecture, with the proportional coefficient Kp set to 0.85, the integral coefficient Ki set to 0.12, and the derivative coefficient Kd set to 0.05. The controller calculates the instantaneous value, cumulative sum, and change slope of the current molten pool size deviation in every 2-millisecond sampling period, and uses the weighted sum of the three as the preliminary feedback correction amount to eliminate the steady-state error caused by environmental fluctuations in the system.

[0072] Furthermore, the system sets the parameters of the proportional-integral-derivative controller based on system identification and optimized tuning. Specifically, under typical conditions of processing titanium alloy powder, i.e., under the reference conditions of laser output power of 350 watts and scanning speed of 800 mm / s, the industrial controller executes a dynamic response experiment of the molten pool width by issuing a step command through the CNC system. During steady-state printing, the target molten pool width is instantaneously stepped from 0.65 mm to 0.75 mm. The discrete data sequence of the actual molten pool width evolution over time is recorded using a synchronous trigger signal. By analyzing the rise time, overshoot, and settling time of the response curve, the physical transfer characteristics of the controlled object are identified, specifically including characteristic elements such as system gain, time constant, and pure time delay.

[0073] Furthermore, the system uses the Ziegler-Nichols tuning rule to calculate the initial values ​​of the control parameters and combines them with a genetic algorithm to perform global optimization. Specifically, based on the ratio of the time constant to the pure time delay obtained by identification, the system calculates the initial values ​​of the proportional coefficient (approximately 0.75), integral coefficient (approximately 0.075), and derivative coefficient (approximately 0.047). Subsequently, the system uses the computing unit of the industrial controller to run the genetic algorithm optimization program. The optimization objective function of the program is defined as minimizing the weighted sum of the settling time and the overshoot. In order to prioritize the suppression of the energy overshoot risk at the thin-walled edge, the weight value of the overshoot is set to 10 and the weight value of the settling time is set to 1. Through 500 generations of evolutionary screening and crossover mutation operations, the optimized combination of control parameters is finally determined to be a proportional coefficient of 0.85, an integral coefficient of 0.12, and a derivative coefficient of 0.05.

[0074] Furthermore, the system incorporates an adaptive parameter correction mechanism for heterogeneous material characteristics to ensure robust power regulation. Specifically, when the processing task is switched to stainless steel or aluminum alloy powder, the controller extracts the thermal diffusivity and laser absorptivity of the current material from the material property database and compares them with the reference value of titanium alloy. By obtaining the square root of the ratio of the current material's thermal diffusivity to the reference value and combining it with the proportional coefficient of the absorptivity change, the controller performs proportional nonlinear correction on the above proportional coefficient, integral coefficient, and differential coefficient. This ensures that the calculation logic of the initial feedback correction amount can automatically adapt to the phase change latent heat requirements and thermal conduction rates of different metal powders, so that the final generated power correction command can maintain a consistent response frequency in multi-material processing scenarios.

[0075] Furthermore, the preliminary feedback correction amount and the feedforward gain coefficient are weighted and synthesized to output the final power correction command. Specifically, the system uses the preliminary feedback correction amount as the main control reference and performs point-by-point multiplication operation with the feedforward gain coefficient obtained above through a multiplication operator. This deeply couples the prediction deviation caused by the geometric position with the measured deviation caused by the actual size of the molten pool, generating a single-precision floating-point control command containing the corrected target power value. This command is sent to the laser's power control board via the EtherCAT real-time bus.

[0076] Furthermore, by adjusting the output power of the laser through the control interface, the instantaneous molten pool feature size converges towards the preset target size. Specifically, after receiving the power correction command, the laser power module adjusts the driving current of its pump source in real time, so that the laser output power dynamically switches within the range of 50W to 500W. If the current instantaneous molten pool feature size is greater than the target value, the system immediately responds by reducing the output power, thereby reducing the heated area of ​​the molten pool. Through this multi-dimensional feature-guided dynamic control, the dynamic deviation between the instantaneous width of the molten pool and the set value is continuously reduced and eventually stabilized within the error range of ±3%, thereby achieving closed-loop quality control of the blade additive manufacturing process.

[0077] Furthermore, the introduction of feedforward gain coefficient for thermal field compensation includes: extracting the closed curve of the current processing layer cross-sectional profile and establishing a heat dissipation gradient prediction model based on the distance field; identifying sharp corner regions with a radius of curvature less than a preset value or narrow regions with a wall thickness less than a threshold in the profile, and performing nonlinear correlation mapping between the feedforward gain coefficient and the heat dissipation gradient of the above regions to preset the heat accumulation peak caused by the obstruction of the heat dissipation path.

[0078] For thin-walled components (such as blade edges) or lattice structures in complex parts, the system pre-scans the geometric support state around the processing path before laser scanning. If the current forming point is located in a region where the heat dissipation cross-section is sharply reduced, the heat dissipation gradient prediction model will predict a decrease in the magnitude of the heat conduction rate. At this time, the feedforward control module adjusts the reference value in the power command in advance to avoid the formation of an uncontrollable thermal field superposition in this local area, thereby maintaining the consistency of the molten pool size. The heat dissipation limitation is upgraded from "post-event compensation" to "pre-event prediction". By using the geometric topology information of the part itself as a priori condition of the control system, the technical problem of overheating and melting at the edge of thin-walled parts is solved from the physical source.

[0079] To address the unsteady thermal field caused by limited heat dissipation in thin-walled, high-curvature regions, a feedforward gain compensation mechanism based on edge geometric distance was introduced. This mechanism can predictively reduce the energy input, effectively suppressing overheating and spheroidization defects of molten metal droplets at the component edges. By combining a weighted synthesis of a PID feedback algorithm and feedforward control, rapid and stable adjustment of power output was achieved, ensuring consistent metallurgical quality across complex feature areas and the solid region of the component.

[0080] Furthermore, the process of generating the power correction command also includes: acquiring the physical property parameters of the metal powder in the current processing layer, including the average particle size of the metal powder, the loose packing density of the powder bed, and the surface tension coefficient of the metal powder in the molten state; establishing a molten pool evolution model based on multiphase flow dynamics, identifying the coupling influence factors of capillary force and metal vapor backflow pressure on the instantaneous molten pool width; and performing nonlinear quadratic correction on the feedforward gain coefficient based on the local thermal conductivity distribution of the metal powder bed in the current processing area, combined with the latent heat demand of the phase change during the transformation of the metal powder from solid to liquid phase, to compensate for the local heat accumulation fluctuations caused by uneven powder particle size, and suppressing the spheroidization tendency and spattering loss of metal droplets at the thin-wall edge by adjusting the laser power output threshold.

[0081] Furthermore, the multiphase flow dynamics molten pool evolution model is constructed based on the competition mechanism between capillary force and metal vapor backflow pressure. Specifically, the capillary force is determined by the product of the surface tension coefficient of the liquid metal and the average curvature of the molten pool surface. The average curvature is approximated by performing a second derivative operation on the molten pool profile curve. This force tends to cause the molten pool surface to contract and reduce the width of the molten pool. The metal vapor backflow pressure is estimated by the saturated vapor pressure obtained by mapping the molten pool surface temperature and enthalpy of vaporization under an exponential function. The surface temperature is obtained by inverting the infrared image of the second sub-beam. The backflow pressure acts on the molten pool surface, generating downward pressure and inducing the molten pool to concave and expand laterally. The system defines a coupling influence factor between 0 and 1 by calculating the proportion of vapor backflow pressure in the total force. When the factor is close to 0, it indicates that the capillary force is dominant and the molten pool is in a stable state. When the factor is close to 1, it indicates that the backflow pressure is too high and the molten pool is prone to splashing instability.

[0082] Furthermore, the system utilizes a deep multilayer perceptron network to perform rapid approximate reasoning on the aforementioned complex physical mechanisms to meet the real-time control requirements of 2-5 milliseconds. Specifically, the multilayer perceptron network comprises one input layer, three fully connected hidden layers, and one output layer. The number of neurons in the hidden layers are arranged sequentially as 128, 64, and 32, respectively. The interlayer activation function adopts the Leaky-ReLU function with a slope of 0.1 to capture the nonlinear characteristics of melt flow. The training data of this network comes from 1000 sets of simulation results generated by computational fluid dynamics software. The simulation conditions cover a combination of powder particle size of 20 to 50 micrometers, surface tension coefficient of 1.5 to 2.0 Newtons per meter, laser power of 200 to 500 watts, and scanning speed of 600 to 1500 millimeters per second. The true values ​​of the coupling influence factors are calculated by extracting the melt pool width and stress distribution under the simulation steady state as labels. The Adam optimizer is used to complete the iterative training of the weight matrix in offline mode to ensure that the network can quickly output the specific values ​​of the current coupling influence factors based on the real-time input powder parameters and laser parameters.

[0083] Furthermore, the system performs nonlinear quadratic correction logic on the feedforward gain coefficient based on the coupling influence factor. Specifically, when the coupling influence factor is less than 0.3, the molten pool is determined to be in the absolutely stable region and the correction increment is set to 0. When the coupling influence factor is between 0.3 and 0.7, the molten pool is determined to have entered the critical unstable region. The system calculates a correction increment between 0 and -0.1 online based on the deviation ratio of the factor value from 0.3. When the coupling influence factor is greater than or equal to 0.7, the molten pool is determined to be in the extremely unstable region. The system sets the correction increment within the deep attenuation range of -0.1 to -0.25. The final feedforward gain coefficient is determined by multiplying the initial gain coefficient obtained from the table with the correction increment plus 1. Through this quadratic correction based on the physical logic of multiphase flow, the system can dynamically adjust the laser power output threshold, thereby effectively suppressing the spheroidization tendency and spattering loss of metal droplets at the thin-walled edge of the blade while compensating for the thermal accumulation fluctuations caused by uneven powder particle size.

[0084] Specifically, before the laser additive manufacturing task of the turbine metal blades is initiated, the central processing unit of the industrial controller first extracts the physical property parameters of the current batch of metal powder from the material management database. These parameters exist in the form of a 512×1 dimension structured feature tensor, including the experimentally determined D50 value with an average particle size of 35 micrometers and a loose packing density of 4.5 g / cm³. 3 The density value and the scalar value of the surface tension coefficient of 1.8 N / m in the molten state at 1600 degrees Celsius were cached in the static random access memory of the local control system as boundary constraints for the melt pool evolution analysis.

[0085] Furthermore, a deep multilayer perceptron network was constructed to perform melt pool evolution model analysis based on multiphase flow dynamics and identify coupling influencing factors. The perceptron network consists of an input layer, three fully connected hidden layers, and an output layer. The number of neurons in the hidden layers were set to 128, 64, and 32 respectively. The interlayer connections used a Leaky-ReLU activation function with a slope of 0.1 to handle the complex gradient negative bias characteristics in melt flow. During the training phase, the model used the Adam optimizer to update the weights by minimizing the mean square error between the predictor and the true value of the discrete element fluid simulation. The learning rate parameter was fixed at 0.0005, and the processing depth of each batch was set to 64 sets of data.

[0086] Specifically, the sensor network receives a multidimensional vector containing the aforementioned powder physical property parameters, the current laser scanning speed of 1200 mm / s, and the instantaneous molten pool width of 0.65 mm output by the previous module as input data. Through nonlinear mapping processing of the weight matrix within the network, a coupling influence factor in the range of 0 to 1 is generated at the output. This factor quantifies the competitive relationship between the surface shrinkage trend of the molten pool caused by capillary force under high-speed laser scanning and the back pressure generated by the vaporization of molten metal. In the calculation logic, the input physical property parameters are scaled to a unit interval using the Min-Max normalization method. By subtracting the minimum value of the powder batch statistics from the original parameters and dividing by the range value, it is ensured that the output coupling influence factor can truly reflect the dynamic instability risk of the molten pool morphology.

[0087] Furthermore, the system utilizes a preset latent heat compensation operator to perform nonlinear secondary correction on the feedforward gain coefficient. Specifically, the system acquires the local thermal conductivity distribution cloud map of the metal powder bed in the current scanning area in real time. By identifying the region where the local thermal conductivity drops from 20 W / m·K to 5 W / m·K due to the porosity of the powder particles, and combining the 300 kJ / kg latent heat energy consumption required for the metal material to change from solid to liquid phase, the secondary correction increment is calculated. Softmax normalization logic is used to assign specific weights to thermal conductivity fluctuations and latent heat demand. The weight values ​​are obtained by mapping the energy proportion index of each indicator in the current thermal field distribution, thereby further correcting the gain amplitude by 15% to 25% on the basis of the original feedforward gain.

[0088] Specifically, the revised instructions suppress spheroidization tendency and spattering at the thin-walled edge by adjusting the laser power output threshold. The system sends control signals in real time through the field-programmable gate array logic unit of the industrial controller to dynamically lock the laser power output hard threshold within a narrow band range of 350W to 420W. When the uneven distribution of powder particle size is detected to cause local heat accumulation fluctuations exceeding 10%, the FPGA controller forcibly reduces the driving current of the pump source to keep the wetting angle of the molten metal droplets at the thin-walled edge below 90%, thereby preventing the molten metal from shrinking back and forming spheroidized particles due to excessive surface tension, and ensuring that the forming density of the blade edge reaches more than 99.8%.

[0089] By deeply modeling and correcting the powder properties, precise control of laser energy at the complex edges of turbine blades was achieved under the condition of uneven micromaterial properties, which effectively improved the melting stability and geometric fidelity in the additive manufacturing process.

[0090] This system achieves closed-loop control of laser power through four core modules: The smelting coaxial sensing module, responsible for signal separation and dynamic adaptation. It uses dichroic elements to separate the processing laser and radiation signals in the physical optical path and adjusts the sensor exposure in real time based on the component's geometric features to ensure high signal-to-noise ratio data acquisition. The spatter noise reduction module, responsible for image purification, uses spatial filtering and adaptive algorithms, combined with a geometric mask constructed from the digital contour of the section to be processed, to accurately filter out powder spatter and outlier noise outside the processing area, extracting a clean molten pool image. The molten pool solid-liquid analysis module, responsible for size restoration, identifies the liquid-solid phase boundary and, combined with motion vectors and normal tilt angles, corrects projection distortion, converting visual pixels into orthogonal physical melt width dimensions that accurately reflect the smelting state. The forming thermal field control module, responsible for control decisions, dynamically generates power correction commands and adjusts the laser output by comparing dimensional deviations and introducing edge heat dissipation feedforward gain, achieving precise closed-loop control of smelting heat input. It is not only applicable to the processing of complex thin-walled parts in the aerospace field, but can also be extended to other high-precision laser thermal processing fields. Through the closed-loop flow of multimodal data and the deep integration of physical laws, it provides a complete technical means to achieve the standardization and intelligence of the additive manufacturing process.

[0091] Example 2: This embodiment is set in the scenario of printing internal cooling channels in hollow blades of aero-engines. It is a specific implementation method for predicting and controlling the distribution of multi-layer accumulated thermal field and residual stress, aiming to solve the problem of thermal stress deformation generated in thin-walled sandwich structures during long-term laser scanning.

[0092] Furthermore, the system performs multi-layer cumulative thermal field integration calculations based on discrete time steps. Specifically, the processing unit of the industrial controller first receives a normalized sequence of molten pool feature images for 50 consecutive scan cycles from the front-end image processing module. The time span corresponding to each frame is 2 milliseconds. The system extracts the average grayscale value of the pixels in the core region of the molten pool in each frame and converts it into instantaneous heat flux scalar data. Then, it performs time integration processing on the heat flux scalar using a sliding observation window with a length of 100 milliseconds. The specific calculation logic is to integrate the heat flux at each sampling time. The numerical value is multiplied by a discrete sampling interval of 0.2 milliseconds, and the total input heat energy at a specific coordinate point in the current layer is accumulated. At the same time, the system dynamically calls a nonlinear heat dissipation attenuation coefficient based on the local wall thickness parameter of the current blade, and subtracts the energy loss value determined by the heat dissipation attenuation coefficient from the total input heat energy, thereby obtaining the residual heat estimate of the coordinate point when printing starts in the next layer. This estimate is processed by Min-Max normalization logic to map the physical temperature from 0 to 1450 degrees Celsius into normalized heat load components between 0 and 1 and output to the stress evaluation model.

[0093] Furthermore, the system utilizes a stress state evaluation model incorporating a long short-term memory network architecture to perform in-depth analysis of the aforementioned time integral results. Specifically, the model consists of an input layer, two hidden layers, and a fully connected output layer. Each hidden layer is configured with 256 recurrent neuron units, and the input layer receives a normalized thermal load time series with a dimension of 50×1. The model uses forget gates, input gates, and output gates in the hidden layers to model the long-distance temporal correlation in the thermal field evolution process. By performing nonlinear mapping in the computational kernel of the graphics processing unit, the thermal field features accumulated over time are transformed into an abstract feature vector representing the residual stress level of the current layer. During the training phase, the model employs the Adam optimizer and iterates the weights by minimizing the mean square error between the predicted stress value and the true value measured by the synchrotron radiation diffraction experiment. The learning rate is set to 0.001, and the sample batch size for each gradient update is set to 32 groups.

[0094] Specifically, the stress state evaluation model uses Softmax processing logic to perform white-box assessment of the output stress risk. Specifically, it calculates the natural exponential mapping ratio of the original components generated by the three neurons in the output layer to obtain three normalized probability values ​​representing safe operation, stress concentration, and crack risk, respectively. When the probability value representing stress concentration exceeds 0.75, the system automatically retrieves the corresponding power compensation strategy through the logic layer, calculates a power reduction factor between 0.8 and 0.95, and uses a multiplication operator to perform weighted synthesis on the preliminary feedback correction amount generated by the aforementioned module. The calculated power correction command is sent to the laser controller through the EtherCAT bus, so that the laser output power is preventively reduced according to the historical thermal accumulation state.

[0095] This embodiment introduces a discrete-time integral algorithm and a recurrent neural network model to quantitatively describe the dynamic evolution of multi-layer thermal fields during the additive manufacturing process of aerospace blades, effectively solving the problem of geometric distortion caused by heat accumulation in complex hollow structures. This implementation method, through precise integration of thermal input over time and deep extraction of nonlinear stress characteristics, enables the control system to predict future layer processing risks, significantly improving the manufacturing yield of aerospace-grade high-performance components.

[0096] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A real-time closed-loop control method for 3D printing laser power, characterized in that, The method includes: The radiation signal generated by the processing laser and the melting of metal powder is separated by a dichroic reflective element embedded in the laser coaxial optical path. The band feature signal containing the three-dimensional morphology of the molten pool and the thermal gradient distribution is extracted from the radiation signal, and the sensor is dynamically compensated for photosensitivity based on the geometric features of the metal component surface. Spatial filtering and adaptive threshold operators are used to process the feature signal of the band. The geometric boundary constraints of the profile of the section to be processed are referenced to filter out the shadows formed by metal powder splashes and high brightness noise interference, so as to obtain a normalized molten pool feature image. The outermost liquid-solid phase boundary of the molten metal in the molten pool feature image is located. Combining the direction of the instantaneous movement vector and the spatial normal tilt angle of the component surface torsion, the instantaneous molten pool feature size orthogonal to the current forming direction is dynamically calculated. The energy deviation is calculated by comparing the instantaneous molten pool characteristic size with the target size, and a feedforward gain coefficient is introduced based on the geometric distance from the current forming point to the edge of the component to perform thermal field compensation, thereby generating a power correction command.

2. The method according to claim 1, characterized in that, Extracting the characteristic signal of the band includes: The radiation signal is divided into a first sub-beam and a second sub-beam by a beam-splitting component disposed on the reflected light path of the dichroic reflective element; the first sub-beam is guided into a first narrowband filter to obtain a feature image reflecting the edge contour of the molten pool, while the second sub-beam is guided into a second narrowband filter to obtain an image reflecting the thermal gradient characteristics inside the molten pool; a synchronous trigger signal is used to drive a dual sensing unit to synchronously acquire the first sub-beam and the second sub-beam. The dynamic photosensitivity compensation of the sensor based on the geometric features of the metal component surface includes: acquiring the spatial distortion parameters of the current formed surface; and dynamically adjusting the integration time of the sensing unit based on the reflectivity change caused by the spatial distortion parameters, so that the acquired band feature signal is within a preset linear response range.

3. The method according to claim 1, characterized in that, The step of processing the band feature signal using spatial filtering and an adaptive threshold operator to obtain a normalized molten pool feature image includes: Spatial smoothing filtering is performed on the characteristic signals of each band, and weighted fusion is performed according to the signal-to-noise ratio of each band to obtain a single-channel feature image; the local energy gradient direction in the single-channel feature image is identified, and the complete molten pool core region is extracted by comparing the statistical differences in gradient distribution between the molten pool core region and the metal powder spatter; the pixel grayscale values ​​of the molten pool core region are normalized to generate the normalized molten pool feature image; The geometric boundary constraints of the reference cross-section profile to be processed, filtering out shadows and high-brightness noise interference formed by metal powder splashes, include: acquiring the digital model profile of the cross-section to be processed and mapping it to the imaging coordinate system of the current sensor to construct a dynamic geometric shielding mask; using the dynamic geometric shielding mask to filter out splash shadows and high-brightness outlier noise located outside the geometric boundary and parallel to the component profile.

4. The method according to claim 1, characterized in that, Locating the outermost liquid-solid boundary of the molten metal in the feature image of the molten pool includes: Extract sub-pixel level coordinate points of the edge regions in the normalized molten pool feature image; use a curve fitting algorithm to fit the sub-pixel level coordinate points to construct a closed contour curve representing the outermost liquid-solid phase line of the molten pool. The dynamic calculation of the instantaneous molten pool feature size orthogonal to the current forming direction includes: obtaining the instantaneous movement vector of the laser beam at the current forming point, establishing a spatial follower coordinate system that dynamically rotates and compensates with the laser beam movement; identifying the normal tilt angle between the spatial normal of the component surface at the current forming point and the laser incident direction, and using a cosine compensation algorithm to correct the geometric distortion of the closed contour curve in the longitudinal direction; and calculating, in the spatial follower coordinate system, the maximum chord length in the direction perpendicular to the instantaneous movement vector in the closed contour curve after geometric distortion correction, as the instantaneous molten pool feature size.

5. The method according to claim 1, characterized in that, The thermal field compensation based on the feedforward gain coefficient introduced according to the geometric distance from the current forming point to the edge of the component includes: The system identifies thin-walled, high-curvature regions within a component that are subject to limited heat dissipation, and calculates the minimum geometric distance from the current forming point to the edge of the component. It then retrieves a preset nonlinear compensation table to obtain the feedforward gain coefficient corresponding to the minimum geometric distance. The feedforward gain coefficient decreases nonlinearly as the current forming point enters the thin-walled, high-curvature region, with the deceleration rate increasing as the distance decreases, thus suppressing overheating caused by the accumulation of molten metal droplets at the edge. The generation of the power correction command includes: calculating the deviation between the instantaneous molten pool feature size and the preset target size, and calculating the preliminary feedback correction amount using a proportional-integral-derivative control algorithm; weighting the preliminary feedback correction amount with the feedforward gain coefficient to output the final power correction command; and adjusting the output power of the laser through the control interface to make the instantaneous molten pool feature size converge towards the preset target size.

6. The method according to claim 5, characterized in that, The process of generating power correction instructions also includes: The physical properties of the metal powder in the current processing layer are obtained, including the average particle size of the metal powder, the loose packing density of the powder bed, and the surface tension coefficient of the metal powder in the molten state. A molten pool evolution model based on multiphase flow dynamics is established to identify the coupling influence factors of capillary force and metal vapor backflow pressure on the instantaneous molten pool width. Based on the local thermal conductivity distribution of the metal powder bed in the current processing area and combined with the latent heat demand of phase change during the transformation of the metal powder from solid to liquid phase, the feedforward gain coefficient is corrected to compensate for the local heat accumulation fluctuations caused by uneven powder particle size. The spheroidization tendency and spattering loss of metal droplets at the thin-wall edge are suppressed by adjusting the laser power output threshold.

7. A real-time closed-loop control system for 3D printing laser power, characterized in that, The system is used to perform the method of claim 1, including: The melting coaxial sensing module separates the radiation signals generated by the processing laser and the melting of metal powder using a dichroic reflective element embedded in the laser coaxial optical path. It extracts band feature signals containing the three-dimensional morphology of the molten pool and the thermal gradient distribution from the radiation signals, and performs dynamic photosensitivity compensation on the sensor based on the geometric features of the metal component surface. The spatter noise reduction module processes the band feature signals using spatial filtering and an adaptive threshold operator. Referring to the geometric boundary constraints of the profile to be processed, it filters out shadows and high-brightness noise interference formed by metal powder spatter, obtaining a normalized molten pool feature image. The molten pool solid-liquid analysis module locates the outermost liquid-solid phase boundary of the molten metal in the molten pool feature image. Combining the direction of the instantaneous movement vector and the spatial normal tilt angle of the component surface's conformal torsion, it dynamically calculates the instantaneous molten pool feature size orthogonal to the current forming direction. The forming thermal field control module compares the instantaneous molten pool feature size with the target size to calculate the energy deviation. Based on the geometric distance from the current forming point to the component edge, it introduces a feedforward gain coefficient for thermal field compensation, generating a power correction command.

Images