Metal 3D printing material performance evaluation method based on edge computing
By using edge computing for data initialization, real-time status updates, and dynamic closed-loop control, the problems of cloud latency and insufficient edge computing power in metal 3D printing have been solved. This enables real-time quantitative evaluation of high-dimensional thermodynamic states and active intervention in microstructure, thereby improving the performance of the formed parts.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ZHEJIANG TUOBAO ADDITIVE MANUFACTURING CO LTD
- Filing Date
- 2026-04-07
- Publication Date
- 2026-07-03
AI Technical Summary
Existing technologies in metal 3D printing suffer from problems such as delayed defect detection due to cloud communication latency, delayed thermodynamic calculations due to insufficient edge computing power, and lack of quantitative evaluation. These issues make it difficult to achieve real-time solutions for high-dimensional thermodynamic states and quantitative evaluation of microscopic properties within extremely short physical time windows.
By constructing a data initialization, real-time state update, residual analysis prediction, and dynamic closed-loop control layer based on edge computing, the physical differential equations are reduced in order using cloud nodes to generate low-dimensional manifold parameters. Real-time temperature gradient calculation and microscopic performance prediction are performed at the edge nodes, and feedforward control commands are generated to adjust process parameters.
It achieves rapid evaluation of high-dimensional thermodynamic state in sub-millisecond time, breaking through the limitations of traditional cloud communication latency and insufficient edge computing power, realizing active intervention of microstructure and self-healing of defects, and improving the yield strength and fatigue life of molded parts.
Smart Images
Figure CN122333872A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of intelligent manufacturing and additive manufacturing technology, specifically to a method for evaluating the performance of metal 3D printing materials based on edge computing. Background Technology
[0002] In the current metal additive manufacturing environment, the processing equipment continuously generates high-frequency thermal radiation data during operation, and the metal metallurgical process has extremely weak Markov properties. The physical properties of the current processing area are greatly constrained by the thermal accumulation effect of the preceding processing path. To monitor and evaluate the processing process, existing solutions generally use cloud computing to calculate high-dimensional thermodynamic states or use traditional edge-side artificial intelligence for surface image analysis. Although these solutions have certain processing capabilities in defect qualitative identification, due to their high dependence on cloud communication, network latency often means that the processing layer has already solidified by the time defects are detected.
[0003] Meanwhile, conventional edge-side models cannot embed complex partial differential equations for quantitative evaluation and lack cross-layer thermal accumulation memory, causing thermal boundary conditions to easily diverge after multi-layer stacking. This results in a serious conflict between computing power and real-time performance, with a lagging process and a lack of quantitative indicators, making it difficult to support rapid intervention before microscopic phase transitions within sub-millisecond physical time windows. Therefore, how to achieve real-time solution of high-dimensional thermodynamic states and quantitative evaluation of microscopic properties on the edge side with extremely limited computing power, so as to complete active closed-loop control before material solidification, has become an urgent technical problem to be solved. Summary of the Invention
[0004] The purpose of this invention is to provide a method for evaluating the performance of metal 3D printing materials based on edge computing, and to solve the following technical problems:
[0005] It overcomes the industry bottleneck that material has already solidified by the time defects are discovered due to cloud communication delays, breaks through the limitations of surface morphology detection and the conflict between low computing power and real-time operation of complex thermodynamic calculations on the edge side, and can achieve active intervention within an extremely short physical time window, guiding the microstructure to complete stress release and defect self-healing at the moment of solidification, thereby significantly improving the yield strength and fatigue life of the final molded part.
[0006] The objective of this invention can be achieved through the following technical solutions:
[0007] A method for evaluating the performance of metal 3D printing materials based on edge computing, comprising:
[0008] A data initialization layer is constructed to receive the material property parameters of the target processing object and the process parameters of the processing equipment. The preset physical differential equations are reduced in order using cloud nodes to extract low-dimensional manifold parameters. A reference thermal history manifold is generated based on the low-dimensional manifold parameters and then sent to the edge nodes.
[0009] A real-time state update layer is constructed to obtain the thermal radiation data collected in real time by the processing equipment during the processing process. The thermal radiation data is inversely solved into transient temperature gradient through the edge nodes, and the cumulative thermal input state matrix of the voxel corresponding to the current processing area is updated based on the transient temperature gradient.
[0010] A residual analysis prediction layer is constructed. At the edge node, the transient temperature gradient is compared with the baseline thermal history manifold to calculate the thermal residual. The thermal residual is input into a preset edge-side lightweight model, and the predicted values of the microscopic performance parameters corresponding to the voxel are output.
[0011] A dynamic closed-loop control layer is constructed, and a feedforward control command data packet is generated based on the predicted values of the microscopic performance parameters. The process control status flag for the processing equipment is updated based on the feedforward control command data packet.
[0012] Optionally, cloud nodes can be used to reduce the order of the preset physical differential equations to extract low-dimensional manifold parameters, including:
[0013] Full-scale finite element simulations are run on the cloud node to generate a full spatiotemporal dataset containing the evolution of molten pool temperature under different process parameters. A loss function is constructed based on the full spatiotemporal dataset and the physical differential equation residuals to train a physical information neural network.
[0014] The physical differential equation is reduced in order to extract the low-dimensional manifold parameters, and the reduced thermodynamic parameter evolution matrix and residual compensation network are generated.
[0015] The training process of the physical information neural network includes: taking material property parameters and process parameters as input, using the temperature field of finite element simulation as data label to calculate data-driven loss, transforming the physical differential equation into a partial differential control equation to calculate physical boundary loss, and minimizing the total loss function through the backpropagation algorithm until the model converges.
[0016] The thermodynamic parameter evolution matrix and the residual compensation network are deployed on the edge nodes to construct the edge-side lightweight model.
[0017] Optionally, generating a reference thermal history manifold based on the low-dimensional manifold parameters includes:
[0018] The material properties and process parameters are converted into tensor inputs.
[0019] Based on the tensor input, an ideal spatial distribution of the temperature field at each location is generated for a specific cross section of the target processing object, and the spatial distribution of the temperature field is used as the reference thermal history manifold; wherein, the material properties include at least specific heat capacity and thermal conductivity.
[0020] Optional, real-time state update layer, specifically configured as follows:
[0021] The thermal radiation data of the processing area of the processing equipment is captured in real time using thermal imaging acquisition equipment;
[0022] The thermal radiation data is inversely calculated to obtain the current transient temperature gradient;
[0023] The transient temperature gradient is used to trigger the edge-side lightweight model, and the cumulative thermal input state matrix within the voxel is updated.
[0024] Optionally, the residual analysis prediction layer is configured as follows: The observed transient temperature gradient is compared with the baseline thermal history manifold to calculate the thermal residual; the thermal residual is substituted into the material solidification kinetics formula in the edge-side lightweight model for solution; the implicit key parameters of the voxels are output as predicted values of microscopic performance parameters; among which, the implicit key parameters include the predicted value of microscopic porosity and the predicted probability of residual stress.
[0025] The observed transient temperature gradient is compared with the baseline thermal history manifold to calculate the thermal residual;
[0026] The thermal residual is substituted into the material solidification kinetics formula in the edge-side lightweight model for solution; the implicit key parameters of the voxel are output as predicted values of the microscopic performance parameters;
[0027] The hidden key parameters include the predicted value of microporosity and the predicted probability of residual stress.
[0028] Optionally, a feedforward control command data packet is generated based on the predicted values of the microscopic performance parameters, and the process control status flag for the processing equipment is updated based on the feedforward control command data packet, configured as follows:
[0029] If the predicted probability of residual stress in the predicted value of the microscopic performance parameters is greater than the preset stress limit threshold, then a feedforward control instruction data packet is generated to indicate the adjustment of the process parameters of the next processing path, and the process control status flag corresponding to the anomaly identifier is updated;
[0030] Otherwise, a data packet is generated to indicate that the current state should be maintained, and the process control status flag corresponding to the normal identifier is updated.
[0031] Optionally, after updating the process control status flag, the method further includes:
[0032] The thermal residuals and their corresponding voxel coordinates that cause the predicted residual stress probability to exceed a preset stress exceedance threshold are stored as historical baselines.
[0033] During the lower-level processing, the historical baseline is extracted as the basis for interlayer temperature compensation and feedback compensation calculation is performed.
[0034] Optionally, the method can be applied to a material processing performance evaluation system;
[0035] The process parameters are laser process parameters;
[0036] Its configuration is to perform in-situ thermal state assessment before the material undergoes a microstructure phase transition, in order to output an intervention instruction data package for defect self-healing.
[0037] The beneficial effects of this invention are:
[0038] 1) This invention reduces the order of physical differential equations in the cloud and distributes the model, and the edge nodes perform inverse calculation and prediction of real-time transient temperature gradients. This architecture effectively solves the conflict between complex thermodynamic calculations and low computing power at the edge, breaks through the traditional cloud communication latency, and realizes rapid evaluation of high-dimensional thermodynamic states within sub-millisecond time windows.
[0039] 2) This invention abandons the traditional qualitative image analysis, transforms high-frequency thermal radiation data into transient temperature gradients, and calculates thermal residuals by comparing them with the baseline thermal history manifold; combined with the edge lightweight model, it can quantitatively output implicit key parameters such as micro porosity and residual stress prediction probability, directly addressing the core of quantitative evaluation of microcrystalline structure and mechanical properties.
[0040] 3) In response to the extremely weak Markov property of metallurgical processes, this invention stores the thermal residuals and voxel coordinates that exceed the limits as historical baselines after updating the process state; when performing the next layer of processing, the baseline is extracted for feedback compensation calculation, which successfully closes the three-dimensional thermodynamic causal chain and prevents the thermal boundary conditions from diverging after multiple layers are superimposed, so as to ensure the prediction accuracy.
[0041] 4) This invention dynamically generates feedforward control commands based on the predicted values of microscopic performance parameters, and adjusts the process parameters of the next processing path before the material undergoes a microstructure phase change; this mechanism seizes the extremely short physical time window before the material solidifies, upgrades the purely passive monitoring to actively curb the proliferation of defects, and guides the microstructure to complete stress release and defect self-healing in an instant. Attached Figure Description
[0042] The invention will now be further described with reference to the accompanying drawings.
[0043] Figure 1This is a flowchart illustrating a method for evaluating the performance of metal 3D printing materials based on edge computing, as provided in an embodiment of this application. Detailed Implementation
[0044] 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.
[0045] Please see Figure 1 A method for evaluating the performance of metal 3D printing materials based on edge computing, comprising:
[0046] A data initialization layer is constructed to receive the material property parameters of the target processing object and the process parameters of the processing equipment. The cloud node is used to reduce the order of the preset physical differential equation to extract low-dimensional manifold parameters. Based on the low-dimensional manifold parameters, a reference thermal history manifold is generated and sent to the edge node.
[0047] A real-time state update layer is constructed to obtain the thermal radiation data collected in real time by the processing equipment during the processing process. The thermal radiation data is inversely solved into transient temperature gradient through edge nodes, and the cumulative thermal input state matrix of the voxel corresponding to the current processing area is updated based on the transient temperature gradient.
[0048] A residual analysis prediction layer is constructed, which compares the transient temperature gradient with the baseline thermal history manifold at the edge nodes to calculate the thermal residual. The thermal residual is input into the preset edge-side lightweight model and outputs the predicted values of the microscopic performance parameters corresponding to the voxels. A dynamic closed-loop control layer is constructed, which generates a feedforward control command data package based on the predicted values of the microscopic performance parameters and updates the process control status flags for the processing equipment based on the feedforward control command data package.
[0049] This embodiment provides a global architecture and execution mechanism for a metal 3D printing material performance evaluation method based on edge computing. Specifically, before printing, the system performs a computationally complex task by cloud nodes, reduces the order of the complex strongly spatiotemporally coupled thermodynamic system, and sends a lightweight reference thermal history manifold to the edge nodes.
[0050] During the printing process, edge nodes no longer process isolated static two-dimensional images, but instead convert high-frequency thermal radiation data into transient temperature gradients to update the cumulative thermal input state matrix within the current three-dimensional voxel.
[0051] Furthermore, the edge node calculates the thermal residual between the current observation state and the cloud baseline state, and uses a lightweight model to directly map the predicted value of the microscopic performance. Based on this, it issues feedforward control commands on the edge side. If, during this process, the edge node fails to fully acquire the thermal radiation data of a certain microsecond-level time slice due to sensor jitter or other reasons, the system will activate the anomaly fallback mechanism, automatically calling the historical state matrix of the previous adjacent voxel and the smooth interpolation algorithm for estimation and supplementation, so as to ensure that the thermodynamic calculation is uninterrupted.
[0052] For example, in the 3D printing process of titanium alloy turbine blades in the aerospace field, the cloud pre-calculates the first... Standard temperature field distribution matrix of the layer under ideal laser power And distribute it to the edge gateway bypassing the printer;
[0053] During printing, the edge gateway receives thermal imager data in real time and calculates the transient temperature field matrix of the current molten pool. The thermal residual matrix obtained by subtracting the two is... When an abnormal peak occurs, the edge gateway immediately generates instructions to adjust the laser status;
[0054] The purpose of this mechanism is to overcome the industry bottleneck caused by cloud communication delays, which means that defects are already solidified by the time they are discovered. It enables high-dimensional thermodynamic state solutions and rapid intervention within a millisecond-level physical time window with extremely limited computing power.
[0055] In a preferred embodiment of the present invention, the cloud node is used to perform order reduction processing on the preset physical differential equation to extract low-dimensional manifold parameters, including: running a full-scale finite element simulation on the cloud node to generate a full spatiotemporal dataset containing the evolution of the melt pool temperature under different process parameters, and constructing a loss function based on the full spatiotemporal dataset and the residual of the physical differential equation to train a physical information neural network.
[0056] The physical differential equation is reduced in order to extract the low-dimensional manifold parameters, and the reduced thermodynamic parameter evolution matrix and residual compensation network are generated.
[0057] The training process of the physical information neural network includes: taking material property parameters and process parameters as input, using the temperature field of finite element simulation as data label to calculate data-driven loss, transforming the physical differential equation into a partial differential control equation to calculate physical boundary loss, and minimizing the total loss function through the backpropagation algorithm until the model converges.
[0058] The thermodynamic parameter evolution matrix and the residual compensation network are deployed on the edge nodes to construct the edge-side lightweight model.
[0059] This embodiment provides a construction and deployment step for a cloud-based lightweight proxy model with physical constraints. Traditional conventional edge AI cannot embed complex partial differential equations, while finite element analysis with thermodynamic computing capabilities cannot run in real time on the edge with low computing power. In order to overcome this absolute conflict between computing power and effectiveness, the system specifically uses high-performance computing power in the cloud to run full-scale finite element simulation and physical information neural network.
[0060] Model reduction techniques are used to reduce the dimensionality of equations such as fluid equations and high-dimensional heat conduction equations; it is assumed that the original physical system possesses... The degrees of freedom are mapped to a low-dimensional manifold space of a preset dimension through singular value decomposition or autoencoder, generating a tiny thermodynamic parameter evolution matrix and residual compensation network, which are then distributed to the edge nodes.
[0061] If, during the cloud-based model reduction process, a certain combination of extreme process parameters is found to cause the manifold space to not converge, the model will be rejected and an abnormal warning will be pushed to the operator terminal indicating that the process parameters exceed the physically solvable range. For example, in the turbine blade printing scenario mentioned above, the cloud compresses the solidification process simulation, which takes several hours, into an evolution matrix derivation model that occupies only a few megabytes of memory and deploys it on the edge industrial personal computer of the printer.
[0062] The purpose of this step is to cleverly transform the computationally complex task of solving partial differential equations into extremely lightweight matrix lookup and residual interpolation calculations on the edge side, which takes less than a millisecond and preserves the causality of physics.
[0063] In a preferred embodiment of the present invention, generating a reference thermal history manifold based on low-dimensional manifold parameters includes: converting material property parameters and process parameters into tensor inputs; generating a spatial distribution of temperature field at each location under ideal conditions for a specific cross section of the target processing object based on the tensor inputs, and using the spatial distribution of temperature field as the reference thermal history manifold; wherein the material property parameters include at least specific heat capacity and thermal conductivity.
[0064] This embodiment provides an initialization mechanism for domain knowledge graphing and structured modeling; specifically, before printing, the system extracts the core physical property parameters and process parameters of the material, and packages them into a multi-dimensional input tensor. ;
[0065] The system will use this tensor The input is fed into the cloud-based reduced-order model. Along the Z-axis slicing direction, for each specific two-dimensional cross-section of the target processing object, the spatial distribution of the temperature field under ideal conditions is derived in a forward direction and solidified into a reference thermal history manifold. If the input combination of physical property parameters is missing in the cloud-based material property library, such as the use of a special alloy powder with a completely new ratio, the system will block the generation of the manifold and require the user to manually input key thermodynamic experimental data such as the solidus / liquidus temperature of the material as backup compensation data.
[0066] For example, for Ti6Al4V titanium alloy, the system measures its specific heat capacity ( ) and thermal conductivity ( The value, along with the current scan path strategy, is converted into a tensor to calculate the turbine blade's first... The ideal thermal history curve of the leaf root region during defect-free solidification; the purpose of this mechanism is to provide a comparative benchmark with rigorous physical meaning for real-time monitoring of the edge side, overcoming the subjective limitations of traditional schemes that rely solely on experience to set image grayscale thresholds.
[0067] In a preferred embodiment of the present invention, the real-time state update layer is specifically configured as follows: using a thermal imaging acquisition device to capture thermal radiation data of the processing area of the processing equipment in real time; inversely solving the thermal radiation data into the current transient temperature gradient; using the transient temperature gradient to trigger the edge-side lightweight model and update the cumulative thermal input state matrix within the voxel.
[0068] This embodiment provides a reverse evolution step for dynamic data stream state update; the metal metallurgical process has extremely weak Markov property, that is, the physical properties of the current region are not only affected by the laser state at this time, but also constrained by the heat accumulation effect brought about by the previous processing path.
[0069] Specifically, after receiving the molten pool plume and thermal radiation data captured by the high-frequency thermal imager, the edge node does not perform any conventional image target detection, such as classifying splashes or unfused parts, but directly extracts them as a two-dimensional optical intensity matrix.
[0070] The system incorporates a temperature measurement function based on Planck's radiation law, converting the grayscale values of the two-dimensional optical intensity matrix into a two-dimensional temperature field distribution on the molten pool surface; further, it incorporates empirical formulas for thermal conduction attenuation along the depth direction of the material:
[0071]
[0072] in, For depth The temperature at that location For the two-dimensional surface temperature, Depth in a three-dimensional coordinate system The thermal decay coefficient is specific to the material, and based on this, the two-dimensional temperature field is extended into a three-dimensional temperature scalar field;
[0073] Taking the partial derivatives of this three-dimensional temperature scalar field along the spatial coordinate axes, we can inversely map it into a transient temperature gradient vector in the three-dimensional coordinate system:
[0074]
[0075] in, The transient scalar temperature in the current three-dimensional space; This is a three-dimensional spatial gradient operator; the transient temperature gradient... As an increment, the three-dimensional cumulative temperature gradient state matrix is added to the current voxel. In the case of local pixel saturation caused by optical path contamination in the thermal imaging acquisition device, such as when the pixel grayscale value reaches the upper limit of 255, the edge node will discard the abnormal frame and use the Kalman filter algorithm to infer the temperature gradient of the current voxel based on the state vector of the preceding normal frame; this transient temperature gradient As an increment, the three-dimensional cumulative temperature gradient state matrix is added to the current voxel. middle;
[0076] If the thermal imaging acquisition device causes local pixel data saturation due to optical path contamination, such as when the pixel grayscale value reaches the upper limit of 255, the edge node will discard the abnormal frame and use the Kalman filter algorithm to infer the temperature gradient of the current voxel based on the state vector of the previous normal frame.
[0077] For example, in the printing of thin-walled blade structures, the thermal imager continuously captures the increase in brightness in the region, and the edge nodes solve it as an extremely steep temperature gradient. The cumulative temperature gradient state matrix of the voxel at the thin-walled region is then updated to indicate that an excessive temperature gradient has been generated in the region. The purpose of this step is to transform the apparent light / thermal signal into the underlying physical properties in real time, and to establish a causal relationship across physical levels.
[0078] In a preferred embodiment of the present invention, the residual analysis prediction layer is specifically configured as follows: the observed transient temperature gradient is compared with the baseline thermal history manifold to calculate the thermal residual; the thermal residual is substituted into the material solidification kinetics formula in the edge-side lightweight model for solution; and the implicit key parameters of the voxel are output as predicted values of microscopic performance parameters; wherein, the implicit key parameters include the predicted value of microscopic porosity and the predicted probability of residual stress.
[0079] This embodiment provides a residual analysis mechanism between model prediction and actual observation. Traditional AI can only qualitatively determine the defect morphology and cannot quantitatively assess material properties, such as the decrease in yield strength. Specifically, the edge node subtracts the actual transient temperature gradient matrix extracted in the previous step from the baseline thermal history manifold matrix sent from the cloud in the corresponding coordinate system to obtain the thermal residual matrix, and then flattens the thermal residual matrix into a one-dimensional thermal residual vector. ;
[0080] thermal residual vector The current cooling rate is calculated by differentiating the transient temperature field differences between adjacent time slices. The components are concatenated to construct a one-dimensional input feature tensor. One-dimensional input feature tensor The dimension is , The total number of features is the concatenation of the thermal residual vector and the cooling rate. This represents the transient temperature difference between adjacent time slices. The time slice step size is defined as follows: The input feature tensor is substituted into the material solidification kinetics formula in the edge-side lightweight model for solution; the kinetics formula is specifically represented as a network layer pre-trained in the cloud, combining a single-layer linear regression operator with an activation function, and its tensor multiplication logic is as follows:
[0081]
[0082] in, The thermodynamic mapping weight tensor distributed from the cloud has the following dimensions: , The number of types of implicit key parameters in the output; For bias terms, Use the Sigmoid activation function to ensure that the output value maps to the 0-1 domain; compute the latent key parameter matrix of the output. The first element directly corresponds to the specific percentage prediction value of the current voxel microporosity, and the second element corresponds to the probability value of severe residual stress exceeding the limit.
[0083] If the calculated residual stress probability is in an oscillating state, such as frequently fluctuating around a threshold, the system will introduce a confidence interval. A definitive warning will only be output when the probability average line for three consecutive sampling periods exceeds the upper limit. The preset stress over-limit threshold is set based on the standard tensile test yield strength of the target material. When the predicted residual stress value reaches 80% of the material's yield strength, it is determined to be greater than the stress over-limit threshold. The confidence interval is configured such that the residual stress prediction probability average line fluctuates within a range of ±5%.
[0084] For example, in the above-mentioned blade printing, it was found that the actual temperature gradient of a certain voxel at the edge of the blade root was much lower than that of the reference manifold. A huge thermal residual was calculated. The kinetic formula was used to quickly deduce that the cooling rate of the voxel was too low, which would generate coarse columnar crystals. The specific quantitative index of 92% residual stress exceeding the limit was output. The purpose of this step is to overcome the limitations of appearance morphology detection and realize the objective quantitative evaluation of microcrystalline structure and mechanical properties.
[0085] In a preferred embodiment of the present invention, a feedforward control command data packet is generated based on the predicted values of microscopic performance parameters, and the process control status flag for the processing equipment is updated based on the feedforward control command data packet, configured as follows:
[0086] If the predicted probability of residual stress in the predicted value of microscopic performance parameters is greater than the preset stress limit threshold, a feedforward control instruction data packet is generated to indicate the adjustment of the process parameters of the next processing path, and the process control status flag corresponding to the abnormality is updated; otherwise, a data packet is generated to indicate maintaining the current state, and the process control status flag corresponding to the normal status is updated.
[0087] This embodiment provides a dynamic closed-loop control mechanism based on interpretable action commands; simply detecting problems without intervention will still lead to the waste of expensive metal powder;
[0088] Specifically, edge nodes poll the implicit key parameter matrix in real time, and when determining the probability of residual stress prediction... Strictly exceeding the preset stress limit threshold Immediately upon activation, the feedforward control logic is initiated, and within microseconds, a feedforward control data packet containing register modification instructions is assembled. This data packet instructs the lower-level system to intervene before the next scan line is executed, and simultaneously marks the process control status of the corresponding voxel in the database as abnormal and indicating intervention.
[0089] like If the laser lower unit is in offline protection mode and cannot receive instructions, the edge system will forcibly cut off the enable signal of the powder bed spreading mechanism and execute an emergency stop of the hardware control loop to avoid greater losses.
[0090] For example, when the probability of residual stress exceeding the limit at the leaf root reaches 92%, the edge node instantly sends a feedforward command to forcibly reduce the laser power of the next scan line in that area by 5% and increase the scan speed by 10% to reduce the line energy density and prevent further excessive heat accumulation. The purpose of this step is to utilize the extremely short physical time window before the solidification of the microstructure to transform purely passive monitoring into active suppression of the generation and expansion of defects.
[0091] In a preferred embodiment of the present invention, after updating the process control status flag, the method further includes: storing the thermal residual that causes the residual stress prediction probability to be greater than the preset stress over-limit threshold and its corresponding voxel coordinates as historical baselines; and extracting the historical baselines as the basis for interlayer temperature compensation calculations during the next layer processing.
[0092] This embodiment provides a cross-layer compensation evolution mechanism for the interlayer thermal accumulation effect; although conventional same-layer feedback intervention is effective, if it does not have cross-layer memory, it is very easy to cause thermal boundary condition collapse and divergence after multiple layers are stacked.
[0093] Specifically, after the edge system completes feedforward control, it not only records the intervention action, but also triggers the core pre-intervention data—namely, the excessive thermal residual. The three-dimensional spatial coordinates of the voxel Encapsulated into a dictionary format, it is persistently written to the edge-side memory database as a historical baseline; when a new layer of powder is laid, the processing equipment processes the projected voxel directly above the coordinate system. At that time, among them For the single-layer pavement thickness increment, the edge model actively extracts the historical baseline data and introduces penalty weights for feedback compensation calculation based on the reference thermal history manifold; penalty weights The calculation satisfies the formula:
[0094]
[0095] in, The interlayer thermal conductivity attenuation constant, The characteristic depth of the heat-affected zone, The maximum allowable thermal residual threshold for the material. For the increase in thickness of a single layer of paving, Historical thermal residuals to be extracted; maximum allowable thermal residual threshold for the material. Based on the calibration curves of porosity and corresponding thermal residual determined by tomographic scanning of standard process specimens in the early stage, the critical thermal residual value at which the porosity changes abruptly is extracted as the maximum thermal residual threshold. ;
[0096] The system utilizes the penalty weights Multiply by the preset process parameter baseline of the current layer to reduce the initial preset laser power of this layer; if the memory database triggers the cleanup mechanism due to storage limit, the system will prioritize retaining the historical baseline of high stress gradient and discard the coordinate data of the stable region;
[0097] For example, in the first leaf The coordinates of the layer where excessive heat accumulation occurs. Its abnormal residuals are stored; when the system prints the first... When the same projection position of the layers is reached, the model predicts that there is residual heat in the bottom layer that has not been completely dissipated. It automatically lowers the initial preset laser power baseline of this layer and performs interlayer temperature compensation. The purpose of this step is to close the thermodynamic causal chain in three-dimensional space so that the prediction accuracy will not diverge exponentially with the increase of printing depth.
[0098] In a preferred embodiment of the present invention, the method is applied to a material processing performance evaluation system; the process parameters are laser process parameters; and it is configured to perform in-situ thermal state evaluation before the material undergoes a microstructure phase transition, so as to output an intervention instruction data package for defect self-healing.
[0099] This embodiment provides the overall system integration and unexpected synergistic effect presentation of the aforementioned evaluation method; specifically, the system deeply couples material processing performance evaluation with laser micro-intervention; due to the superposition of physical order reduction model and extremely low latency edge computing, the method can complete high-dimensional physical solution in sub-millisecond time, which is much shorter than the physical time window required for metal powder to undergo micro-martensitic phase transformation;
[0100] If the system fails to successfully send the intervention data packet before the phase transition critical point, it will automatically mark the area as an irreversible defect area and generate a flaw detection suggestion report after printing is completed. For example, for the above-mentioned titanium alloy turbine blade, the system uses microsecond-level instructions to continuously change the local metallographic evolution path of the metal during the printing and evaluation process.
[0101] The purpose of this mechanism is to achieve a paradigm shift, that is, instead of repairing cracks after they have formed, it uses extremely short feedback loops generated by ultra-fast calculations to guide the microstructure to complete stress release and self-healing of defects at the moment of solidification, thereby significantly improving the yield strength and fatigue life of the final molded part.
[0102] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.
Claims
1. A method for metal 3D printing material performance evaluation based on edge computing, characterized in that ,include: A data initialization layer is constructed to receive the material property parameters of the target processing object and the process parameters of the processing equipment. The preset physical differential equations are reduced in order using cloud nodes to extract low-dimensional manifold parameters. A reference thermal history manifold is generated based on the low-dimensional manifold parameters and then sent to the edge nodes. A real-time state update layer is constructed to obtain the thermal radiation data collected in real time by the processing equipment during the processing process. The thermal radiation data is inversely solved into transient temperature gradient through the edge nodes, and the cumulative thermal input state matrix of the voxel corresponding to the current processing area is updated based on the transient temperature gradient. A residual analysis prediction layer is constructed. At the edge node, the transient temperature gradient is compared with the baseline thermal history manifold to calculate the thermal residual. The thermal residual is input into a preset edge-side lightweight model, and the predicted values of the microscopic performance parameters corresponding to the voxel are output. A dynamic closed-loop control layer is constructed, and a feedforward control command data packet is generated based on the predicted values of the microscopic performance parameters. The process control status flag for the processing equipment is updated based on the feedforward control command data packet.
2. The method for evaluating the performance of metal 3D printing materials based on edge computing according to claim 1, characterized in that... The step of using cloud nodes to reduce the order of a preset physical differential equation to extract low-dimensional manifold parameters includes: Full-scale finite element simulations are run on the cloud node to generate a full spatiotemporal dataset containing the evolution of molten pool temperature under different process parameters. A loss function is constructed based on the full spatiotemporal dataset and the physical differential equation residuals to train a physical information neural network. The physical differential equation is reduced in order to extract the low-dimensional manifold parameters, and the reduced thermodynamic parameter evolution matrix and residual compensation network are generated. The training process of the physical information neural network includes: taking material property parameters and process parameters as input, using the temperature field of finite element simulation as data label to calculate data-driven loss, transforming the physical differential equation into a partial differential control equation to calculate physical boundary loss, and minimizing the total loss function through the backpropagation algorithm until the model converges. The thermodynamic parameter evolution matrix and the residual compensation network are deployed on the edge nodes to construct the edge-side lightweight model.
3. The method for evaluating the performance of metal 3D printing materials based on edge computing according to claim 1, characterized in that... The step of generating a reference thermal history manifold based on the low-dimensional manifold parameters includes: The material properties and process parameters are converted into tensor inputs. Based on the tensor input, an ideal temperature field spatial distribution is generated for each location at a specific cross section of the target processing object, and the temperature field spatial distribution is used as the reference thermal history manifold; wherein, the material properties include at least specific heat capacity and thermal conductivity.
4. The method for evaluating the performance of metal 3D printing materials based on edge computing according to claim 1, characterized in that... The real-time state update layer is specifically configured as follows: The thermal radiation data of the processing area of the processing equipment is captured in real time using thermal imaging acquisition equipment; The thermal radiation data is inversely calculated to obtain the current transient temperature gradient; The transient temperature gradient is used to trigger the edge-side lightweight model, and the cumulative thermal input state matrix within the voxel is updated.
5. The method for evaluating the performance of metal 3D printing materials based on edge computing according to claim 1, characterized in that... The residual analysis prediction layer is specifically configured as follows: The observed transient temperature gradient is compared with the baseline thermal history manifold to calculate the thermal residual; the thermal residual is substituted into the material solidification kinetics formula in the edge-side lightweight model for solution; the implicit key parameters of the voxels are output as predicted values of microscopic performance parameters; among which, the implicit key parameters include the predicted value of microscopic porosity and the predicted probability of residual stress.
6. The method for evaluating the performance of metal 3D printing materials based on edge computing according to claim 5, characterized in that... The step of generating a feedforward control command data packet based on the predicted values of the microscopic performance parameters, and updating the process control status flag for the processing equipment based on the feedforward control command data packet, is configured as follows: If the predicted probability of residual stress in the predicted value of the microscopic performance parameters is greater than the preset stress limit threshold, then a feedforward control instruction data packet is generated to indicate the adjustment of the process parameters of the next processing path, and the process control status flag corresponding to the abnormality is updated; otherwise, a data packet is generated to indicate maintaining the current state, and the process control status flag corresponding to the normality is updated.
7. The method for evaluating the performance of metal 3D printing materials based on edge computing according to claim 6, characterized in that... After updating the process control status flag, the method further includes: The thermal residuals and their corresponding voxel coordinates that cause the predicted residual stress probability to exceed a preset stress exceedance threshold are stored as historical baselines. During the lower-level processing, the historical baseline is extracted as the basis for interlayer temperature compensation and feedback compensation calculation is performed.
8. A method for evaluating the performance of metal 3D printing materials based on edge computing according to any one of claims 1-7, characterized in that... The method is applied to a material processing performance evaluation system. The process parameters are laser process parameters; Its configuration is to perform in-situ thermal state assessment before the material undergoes a microstructure phase transition, in order to output an intervention instruction data package for defect self-healing.