Aerospace engine turbine blade additive manufacturing defect correction method and system

Abstract

The present application relates to the field of industrial manufacturing, and provides a defect correction method and system for additive manufacturing of aerospace engine turbine blades. The method comprises: adsorbing quantum dot markers on the surface of metal powder, collecting signal data in the additive manufacturing process to obtain a set of multi-dimensional monitoring characteristic parameters; constructing a graph neural network model, inputting the set of multi-dimensional monitoring characteristic parameters as node features, predicting the spatio-temporal evolution of the molten pool area to obtain a defect probability distribution field; performing multi-objective optimization on the process parameter vector through a Bayesian optimization algorithm to obtain an adaptive control parameter combination, and performing partition control on the manufacturing process; for residual defects of the partition control result, using a reinforcement learning strategy network to select a repair action sequence, inputting the repair action sequence into an execution unit to perform cross-level collaborative processing on the residual defects to obtain a repaired blade structure. The present application realizes intelligent prediction and active correction of defects in additive manufacturing of turbine blades.

Description

Technical Field

[0001] This invention relates to the field of industrial manufacturing technology, and in particular to a method and system for defect correction in additive manufacturing of aerospace engine turbine blades. Background Technology

[0002] Turbine blades are core hot-end components of aero-engines, operating in extremely harsh environments. They must withstand temperatures exceeding 1000°C, centrifugal stress from high-speed rotation, and complex loads from exhaust gas scouring. Additive manufacturing technology, as an advanced manufacturing method, enables the integrated forming of complex internal cooling channels in turbine blades, significantly shortening the manufacturing cycle and reducing material waste.

[0003] However, the rapid heating and cooling cycles, complex molten pool flow behavior, and layer-by-layer accumulation characteristics during additive manufacturing easily lead to various defects such as porosity, cracks, lack of fusion, and compositional segregation, severely affecting the mechanical properties and service reliability of blades. How to monitor and suppress defect formation in real time during additive manufacturing has become a key bottleneck restricting the application of this technology in the aerospace field.

[0004] Existing additive manufacturing quality control methods primarily rely on offline post-manufacturing inspections, such as X-ray CT scans and ultrasonic testing. These methods cannot intervene in the defect formation process in a timely manner, making defects difficult to repair once they occur and resulting in a high scrap rate. Although some studies have attempted to use sensors such as infrared thermal imagers and high-speed cameras for online monitoring, these methods can only acquire surface information and lack effective perception capabilities for defect evolution within the molten pool and subsurface layers. Furthermore, they lack intelligent defect prediction and adaptive process control mechanisms. Therefore, current technologies cannot achieve deep, real-time perception, accurate prediction, and intelligent proactive correction of defects during additive manufacturing, leading to low yield and poor quality stability in turbine blade additive manufacturing. Summary of the Invention

[0005] This invention provides a method and system for defect correction in additive manufacturing of aerospace engine turbine blades, in order to overcome the deficiencies of the prior art.

[0006] This invention provides a method for defect correction in additive manufacturing of aerospace engine turbine blades, comprising:

[0007] S1. Adsorb quantum dot labeling agents onto the surface of metal powder, and collect signal data during the additive manufacturing process through a multi-physics field sensor to obtain a set of multi-dimensional monitoring characteristic parameters.

[0008] S2. Construct a graph neural network model, take the multi-dimensional monitoring feature parameter set as node feature input, and predict the spatiotemporal evolution of the molten pool region through an adversarial training mechanism to obtain the defect probability distribution field.

[0009] S3. Based on the defect probability distribution field, the process parameter vector is optimized through a Bayesian optimization algorithm to obtain an adaptive control parameter combination. The manufacturing process is then controlled in zones according to the adaptive control parameter combination to obtain the zone control result.

[0010] S4. For the residual defects in the partition control results, a reinforcement learning strategy network is used to select a repair action sequence. The repair action sequence is input into the execution unit to perform cross-level collaborative processing on the residual defects to obtain the repaired blade structure.

[0011] According to the present invention, a defect correction method for additive manufacturing of aerospace engine turbine blades, step S1 further includes:

[0012] S11. Core-shell quantum dots are adsorbed onto the surface of nickel-based superalloy powder by electrostatic self-assembly to obtain functionalized labeled powder.

[0013] S12. Collect the bimodal fluorescence spectrum, terahertz transmission spectrum, eddy current conductivity signal, laser ultrasonic echo signal and magneto-optical Kerr signal of the functionalized labeled powder under laser irradiation to obtain a five-dimensional raw monitoring data stream.

[0014] S13. Perform Fourier transform and wavelet multi-scale decomposition on the five-dimensional original monitoring data stream to extract frequency domain energy distribution, phase delay, time-frequency localization features and singular values ​​to obtain a multi-dimensional monitoring feature parameter set.

[0015] According to the present invention, in a method for defect correction in additive manufacturing of aerospace engine turbine blades, in step S11, the core-shell structure quantum dot is a CdSe / ZnS or InP / ZnS material system.

[0016] According to the defect correction method for additive manufacturing of aerospace engine turbine blades provided by the present invention, step S2 further includes:

[0017] S21. Discretize the blade manufacturing space into a voxel grid, and use each voxel as a graph node. Establish spatial adjacency edges and physical causal edges between nodes according to the heat conduction direction, stress propagation path and melt flow trajectory to obtain a spatiotemporal dynamic graph structure.

[0018] S22. Construct an adversarial generative graph attention network, wherein the adversarial generative graph attention network includes a generator, a discriminator, and a physical constraint module;

[0019] S23: Assign the multi-dimensional monitoring feature parameter set to the nodes of the spatiotemporal dynamic graph structure, input it into the generator for multi-head attention aggregation calculation, and obtain preliminary prediction results; the preliminary prediction results include: temperature field prediction, stress field prediction, and preliminary defect distribution.

[0020] S24. Update the network parameters of the adversarial generative graph attention network based on the preliminary prediction results;

[0021] S25. The difficulty of the training samples is adjusted by adopting a course learning strategy. The generator outputs a four-dimensional tensor form of the defect probability distribution field, labeling the porosity probability, crack probability, inclusion probability and segregation probability of each voxel.

[0022] According to the defect correction method for additive manufacturing of aerospace engine turbine blades provided by the present invention, step S24 further includes:

[0023] S241. Input the preliminary defect distribution and the real defect distribution samples into the discriminator respectively, and calculate the authenticity score through multi-layer graph convolution. Based on the authenticity score, use the Wasserstein distance metric to predict the difference between the distribution and the real distribution to obtain the adversarial loss value.

[0024] S242. Substitute the temperature field prediction into the Navier-Stokes equation to calculate the temperature residual, and substitute the stress field prediction into the thermoelastic constitutive equation to calculate the stress residual. Perform norm integration on the temperature residual and the stress residual in the spatial domain and the time domain to obtain the physical constraint loss value.

[0025] S243. The adversarial loss value and the physical constraint loss value are weighted and summed, and a regularization term is added to construct the total loss function;

[0026] S244: Calculate the gradient of the total loss function with respect to the parameters of the generator and the discriminator through the optimizer, update the network parameters of the adversarial generative graph attention network using a learning rate decay strategy, and determine convergence when the rate of change of the total loss function is less than a set threshold.

[0027] According to the defect correction method for additive manufacturing of aerospace engine turbine blades provided by the present invention, step S3 further includes:

[0028] S31. Set the defect suppression degree, forming efficiency, surface roughness and residual stress as optimization objectives, and use the laser power, scanning speed, spot diameter, powder feeding rate, protective gas flow rate, substrate temperature, scanning strategy angle, overlap rate and layer thickness as decision variables to construct a multi-objective optimization problem.

[0029] S32. Construct a Gaussian process surrogate model, and encode the quantum dot fluorescence decay time and terahertz absorption coefficient in the multidimensional monitoring feature parameter set into the prior covariance structure of the physical information kernel function to obtain a surrogate model with physical constraints.

[0030] S33. Based on the multi-objective optimization problem and the surrogate model, the candidate parameter set is used to approximate the posterior distribution of the parameter space through variational inference. The Pareto optimal solution set is output through knowledge gradient balancing sampling and multiple virtual evaluations, and the adaptive control parameter combination is obtained by screening.

[0031] S34. Based on the defect probability distribution field, the scanning path is divided into several sub-regions. For sub-regions where the defect probability exceeds the threshold, corresponding adaptive control parameter combinations are assigned to perform partitioned asynchronous control to obtain partitioned control results.

[0032] According to the present invention, a defect correction method for additive manufacturing of aerospace engine turbine blades, in step S32, the construction of the physical information kernel function includes:

[0033] S321. Define the basic kernel function, express the similarity between two points in the parameter space in the form of radial basis functions, and set the signal variance and multiple feature length scales.

[0034] S322. Extract the quantum dot fluorescence decay time and terahertz absorption coefficient from the set of multidimensional monitoring feature parameters, and construct a physical correction factor. The physical correction factor reflects the physical response characteristics when the process parameter vector deviates from the standard value.

[0035] S323. Multiply the physical correction factor by the basic kernel function to obtain the physical information kernel function, which is used to construct the covariance matrix of the surrogate model of the Gaussian process, thereby obtaining a surrogate model with physical constraints.

[0036] According to the defect correction method for additive manufacturing of aerospace engine turbine blades provided by the present invention, step S33 further includes:

[0037] S331. Based on the multi-objective optimization problem and relying on the surrogate model, determine the range of candidate parameter combinations in the parameter space and obtain the candidate parameter set;

[0038] S332. The variational inference method is used to approximate the posterior distribution of the candidate parameter set, and the expected defect rate mean and 95% confidence interval of multiple candidate parameter combinations are calculated through a surrogate model.

[0039] S333. Explore and utilize the parameter space of function trade-offs through knowledge gradients, and prioritize sampling candidate parameter combinations with the largest gain.

[0040] S334. After multiple virtual evaluations, the candidate parameter combinations and prediction results are iteratively updated, and the Pareto optimal solution set that satisfies the multi-objective optimization constraints is output.

[0041] S335. Based on the aforementioned defect probability distribution field characteristics, select adaptive control parameter combinations from the Pareto optimal solution set that are suitable for defect conditions in different sub-regions.

[0042] According to the defect correction method for additive manufacturing of aerospace engine turbine blades provided by the present invention, step S4 further includes:

[0043] S41. Construct a partially observable Markov decision process model, defining the multidimensional monitoring feature parameter set of the current layer and the multilayers already formed below as the state space, defining local remelting, laser shock strengthening, ultrasonic micro-forging, arc-enhanced deposition and cold spray filling and their sub-parameters as the action space, and defining the weighted combination of defect elimination rate and penalty for introducing new defects as the reward function.

[0044] S42. Establish a deep deterministic policy gradient network, which includes a two-stream Actor network and a Critic network. Input the state space into the convolutional stream and recurrent stream of the two-stream Actor network to process spatial features and temporal dependencies respectively, and output continuous repair parameters. The Q value of the state-action pair is evaluated by the Critic network.

[0045] S43. Load the pre-trained multi-material repair strategy library through the meta-learning module, perform rapid adaptation with few samples based on the multi-dimensional monitoring feature parameter set of the current blade pre-sequence layer, and output the repair action sequence for the residual defect by the dual-stream Actor network.

[0046] S44. Drive the femtosecond laser probe to scan along the path of the residual defect to obtain the defect depth and direction data. Perform multi-stage combined processing on the defect area corresponding to the residual defect according to the repair strategy determined by the repair action sequence to obtain the repaired blade structure.

[0047] The present invention also provides a defect correction system for additive manufacturing of aerospace engine turbine blades, for performing a defect correction method for additive manufacturing of aerospace engine turbine blades as described in any of the preceding claims, comprising:

[0048] Acquisition module: used to collect signal data during the additive manufacturing process of quantum dot labeling agents adsorbed on the surface of metal powder, and obtain a set of multi-dimensional monitoring characteristic parameters;

[0049] Prediction module: It receives a set of multi-dimensional monitoring feature parameters as node feature inputs, and performs spatiotemporal evolution prediction of the molten pool region through an adversarial training mechanism to obtain the defect probability distribution field;

[0050] Optimization module: used to perform multi-objective optimization of the process parameter vector based on the defect probability distribution field using a Bayesian optimization algorithm to obtain an adaptive control parameter combination;

[0051] Execution module: used to perform zoned control of the manufacturing process according to the adaptive control parameter combination to obtain zoned control results; also used to execute the repair action sequence output by the repair module to perform cross-level collaborative processing of residual defects;

[0052] Repair module: Used to select repair action sequences using a reinforcement learning strategy network to address residual defects in the partitioning control results.

[0053] This invention provides a defect correction method and system for additive manufacturing of turbine blades for aerospace engines. First, by adsorbing quantum dot markers onto the surface of metal powder and collecting multi-physics field information such as fluorescence, electromagnetic, and acoustic signals, a multi-dimensional monitoring feature parameter set with multiple feature dimensions is constructed. This enables deep, real-time perception of the evolution of defects inside the molten pool and at the subsurface level during additive manufacturing, overcoming the limitation of traditional monitoring methods that can only acquire surface information. The spatial resolution of defect detection is improved to the nanometer level, and the temporal response speed reaches the millisecond level. Second, the adversarial generative graph attention network constructed in this invention uses the multi-dimensional monitoring feature parameter set as node features and combines it with physical causal edges to establish a spatiotemporal dynamic graph structure. This allows for accurate prediction of the temperature field, stress field, and defect probability distribution field of the next 3-10 layers, with high prediction accuracy, enabling defect intervention. This invention shifts from passive response to proactive prevention, effectively avoiding the problem of difficult repair after defects form. Furthermore, the Bayesian optimization algorithm combined with a physical information kernel function performs multi-objective optimization of the process parameter vector. Through adaptive parameter combination control, it achieves partitioned asynchronous control, reducing defect incidence while maintaining high forming efficiency, improving surface roughness, and lowering residual stress levels, significantly enhancing the quality stability of turbine blade additive manufacturing. Additionally, the reinforcement learning strategy network combined with a meta-learning module introduced in this invention can intelligently select repair action sequences for residual defects and perform cross-level collaborative processing, improving the repair success rate and increasing the yield of turbine blade additive manufacturing compared to traditional methods. This significantly reduces manufacturing costs and cycle time, providing an effective technical solution for high-quality additive manufacturing of aerospace engine turbine blades. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0055] Figure 1A schematic flowchart of a defect correction method for additive manufacturing of aerospace engine turbine blades provided by the present invention;

[0056] Figure 2 This invention provides a schematic diagram of a defect correction system for additive manufacturing of aerospace engine turbine blades. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, embodiments of this invention, and should not be construed as limiting the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention. In the description of this invention, it should be understood that the terminology used is for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0058] The embodiments of the present invention are described below with reference to the figures.

[0059] like Figure 1 As shown, the present invention provides a defect correction method for additive manufacturing of aerospace engine turbine blades, comprising:

[0060] S1. Adsorb quantum dot labeling agents onto the surface of metal powder, and collect signal data during the additive manufacturing process through a multi-physics field sensor to obtain a set of multi-dimensional monitoring characteristic parameters.

[0061] Step S1 further includes:

[0062] S11. Core-shell quantum dots are adsorbed onto the surface of nickel-based high-temperature alloy powder by electrostatic self-assembly to obtain functionalized labeled powder; the core-shell quantum dots are CdSe / ZnS or InP / ZnS material systems.

[0063] In step S11, the present invention first pre-treats the surface of nickel-based high-temperature alloy powder by ultrasonic cleaning with ethanol and deionized water to remove surface oxides and oil stains. Then, a positively charged polyelectrolyte layer is modified on the powder surface, and core-shell structured quantum dots are adsorbed by electrostatic self-assembly. Specifically, the core-shell structured quantum dots adopt a CdSe / ZnS or InP / ZnS material system, with the core particle size controlled at 2-5nm and the shell thickness controlled at 1-3nm.

[0064] The specific adsorption process is as follows: First, quantum dots are dispersed in toluene solvent to prepare a solution with a concentration of 1-5 mg / mL. Pretreated nickel-based superalloy powder is then immersed in the quantum dot solution and stirred at 25°C for 2-4 hours. Electrostatic attraction causes the negatively charged quantum dots to adsorb onto the positively charged powder surface. After adsorption, centrifugation and vacuum drying are performed to control the adsorption density at 10-1. 11 -10 13 pcs / cm 2 Within the specified range, functionalized labeled powders were obtained.

[0065] The quantum dots in the functionalized labeled powder obtained in the end can produce a bimodal fluorescence spectrum with a wavelength range of 500-700 nm under subsequent laser irradiation. The 520 nm peak corresponds to the band-edge emission of the quantum dots, and the 680 nm peak corresponds to the surface state emission. When nanoscale pores or dislocations appear around the powder, the fluorescence lifetime of the quantum dots is shortened from the original 50-80 ns to 20-40 ns, and the fluorescence intensity decreases by more than 40%.

[0066] S12. Collect the bimodal fluorescence spectrum, terahertz transmission spectrum, eddy current conductivity signal, laser ultrasonic echo signal and magneto-optical Kerr signal of the functionalized labeled powder under laser irradiation to obtain a five-dimensional raw monitoring data stream.

[0067] In step S12, the present invention deploys a multi-physics sensing system within the forming cavity of the additive manufacturing equipment to simultaneously acquire various physical response signals of the functionalized labeled powder under laser irradiation. Within the multi-physics sensing system, a hyperspectral camera array is deployed to acquire bimodal fluorescence spectral signals with a spectral resolution of 2 nm. The intensity ratio of the two characteristic peaks at 520 nm and 680 nm and the fluorescence lifetime decay curve reflect the defect formation tendency in the molten pool region. A terahertz time-domain spectroscopy system emits terahertz waves with a frequency range of 0.1-4 THz that penetrate a depth of 3-5 mm below the molten pool. After receiving the transmitted signal, the terahertz absorption coefficient is calculated. This coefficient is directly related to the integrity of the material's crystal lattice structure; the absorption coefficient is affected when microcracks or grain boundary separation occur. The conductivity increases by 15-30%; while the eddy current sensor array monitors local conductivity changes at an excitation frequency of 10MHz. Due to the uneven distribution of alloying elements caused by component segregation, the conductivity fluctuates by 5-10%; for the laser ultrasonic detection unit, it emits a laser with a pulse width of 5ns to excite longitudinal and transverse waves in the megahertz band on the material surface. The elastic modulus is calculated by measuring the difference in wave velocity. The presence of pores will reduce the elastic modulus by 20-40%; the magneto-optical Kerr effect measurement device monitors the magnetization intensity change of the ferromagnetic phase. During the phase transition, the magnetization intensity exhibits a sudden jump. The sampling frequency of all the above sensors is uniformly set to 1-5kHz. The single-layer scanning process generates approximately 50GB of five-dimensional raw monitoring data stream, including fluorescence spectral matrix, terahertz time-domain waveform, real and imaginary parts of eddy current impedance, ultrasonic echo time series, and magneto-optical Kerr rotation angle series, ultimately forming a five-dimensional raw monitoring data stream.

[0068] S13. Perform Fourier transform and wavelet multi-scale decomposition on the five-dimensional original monitoring data stream to extract frequency domain energy distribution, phase delay, time-frequency localization features and singular values ​​to obtain a multi-dimensional monitoring feature parameter set.

[0069] In step S13, the present invention extracts frequency domain and time-frequency domain features from the five-dimensional original monitoring data stream. Specifically, for the bimodal fluorescence spectral data, a fast Fourier transform is performed on the intensity time series of the 520nm and 680nm peaks to extract the power spectral density distribution in the 0-500Hz frequency band as the frequency domain energy distribution feature. The low-frequency component (0-50Hz) in the spectrum reflects the overall temperature fluctuation of the molten pool, while the high-frequency component (200-500Hz) reflects abrupt changes in the local temperature gradient. The terahertz transmission signal is decomposed using wavelet multi-scale decomposition, with the Daubechies wavelet selected as the mother wavelet and a decomposition layer of 5 layers. Time-frequency localization features are extracted from the decomposition coefficients. The detail coefficients of layers 1-2 reflect surface defects in the material, while the detail coefficients of layers 3-5 reflect deep defects. For the eddy current conductivity signal, the present invention calculates the phase delay angle between the real and imaginary parts. A phase delay fluctuating between 30-60 degrees indicates a tendency for component segregation. A Hankel matrix was constructed from the laser-ultrasonic echo signal, and singular value decomposition was performed. The top 20 largest singular values ​​were extracted as characteristic parameters. A sudden drop in singular values ​​indicates the presence of a reflection interface, i.e., an internal defect, along the wave propagation path. The instantaneous amplitude and frequency were extracted from the magneto-optical Kerr signal using Hilbert transform.

[0070] Finally, this invention combines the frequency domain energy distribution, phase delay, time-frequency localization features, singular values ​​and other statistical features extracted from the five-dimensional original monitoring data stream to obtain a multi-dimensional monitoring feature parameter set, with each voxel position corresponding to a feature vector.

[0071] S2. Construct a graph neural network model, use the multidimensional monitoring feature parameter set as node feature input, and predict the spatiotemporal evolution of the molten pool region through an adversarial training mechanism to obtain the defect probability distribution field.

[0072] Step S2 further includes:

[0073] S21. Discretize the blade manufacturing space into a voxel grid. Using each voxel as a graph node, establish spatial adjacency edges and physical causal edges between nodes based on the heat conduction direction, stress propagation path, and melt flow trajectory to obtain a spatiotemporal dynamic graph structure.

[0074] In step S21, the present invention discretizes the manufacturing space based on the blade CAD model, setting a grid spacing of 0.1 mm in the x, y, and z directions, dividing the entire blade space into approximately 3 million voxel units, with each voxel serving as a node in the graph structure. When establishing node connections, in addition to establishing 6-neighborhood spatial adjacency edges (i.e., each voxel connects to its six adjacent voxels above, below, left, right, front, and back), the present invention also determines the dominant heat flow direction based on Fourier's law of heat conduction, establishing long-range connection edges in directions with large temperature gradients, with connection distances reaching 5-10 voxel intervals; based on the stress propagation path, establishing physical causal edges along the principal stress direction, with connection distances of 3-8 voxel intervals; and determining the flow trajectory based on the Marangoni effect and gravity of melt flow, establishing directed edges along the flow direction. In the time dimension, the present invention treats each layer scan as a time step, establishing temporal connection edges between nodes at the same spatial location in adjacent time steps. The final spatiotemporal dynamic graph structure contains multiple nodes, each with an average connectivity of 15-25. It includes both local spatial neighborhood information and long-distance physical causal relationships and temporal evolution relationships.

[0075] S22. Construct an adversarial generative graph attention network, which includes a generator, a discriminator, and a physical constraint module.

[0076] This invention constructs an adversarial generative graph attention network, which comprises three modules: a generator, a discriminator, and a physical constraint module. The generator employs a multi-layer stacked graph attention layer structure with 8-16 attention heads, each equipped with an independent learnable weight matrix. , , These are used to calculate the query vector, key vector, and value vector, respectively. For nodes... Its query vector ,in For nodes eigenvectors for nodes; Neighboring nodes Its key vector value vector ,in For nodes The feature vector. The attention coefficient is then calculated first. and The dot product yields the original relevance score, which is then processed by the LeakyReLU activation function and softmax normalization to obtain the attention weights. final node The updated features are calculated by weighted summation as follows: After parallel computation by 8-16 attention heads, the output features are concatenated or averaged. Among the attention heads, the first 5 specifically learn the heat conduction and convection heat transfer interaction modes of the temperature field, the 6th-10th attention heads learn the elastoplastic coupling mode of the stress field, and the 11th-16th attention heads learn the diffusion and segregation evolution modes of alloying elements. The discriminator employs a 5-layer graph convolutional network structure, with each layer containing graph convolution operations and batch normalization. Finally, a truth score between 0 and 1 is output through global average pooling and fully connected layers. The physical constraint module embeds the discretized form of the Navier-Stokes equations and the stress-strain relationship of the thermoelastic constitutive equations.

[0077] S23: Assign the multi-dimensional monitoring feature parameter set to the nodes of the spatiotemporal dynamic graph structure, input it into the generator for multi-head attention aggregation calculation, and obtain preliminary prediction results; the preliminary prediction results include: temperature field prediction, stress field prediction and preliminary defect distribution.

[0078] Furthermore, this invention assigns the 150-dimensional feature vector corresponding to each voxel in the multidimensional monitoring feature parameter set to the corresponding node of the spatiotemporal dynamic graph structure. The assigned graph structure is then input into the generator, and the graph data is processed in mini-batch mode, with each batch containing 5000-10000 nodes.

[0079] During the forward propagation of the generator, the first graph attention layer receives the initial node features and aggregates first-order neighborhood information through a multi-head attention mechanism; the second graph attention layer aggregates second-order neighborhood information based on the updated node features; and so on. After 4-6 graph attention layers, the feature vector of each node incorporates the information of all relevant nodes within a 4-6 hop range around it.

[0080] The generator's output layer contains three branches: a temperature prediction branch that outputs the temperature values ​​for each voxel in the next 3-10 layers, and a stress prediction branch that outputs the six stress components for each voxel in the next 3-10 layers. The defect prediction branch outputs the probability values ​​for four types of defects—porosity, cracks, inclusions, and segregation—for each voxel. The temperature field prediction is a four-dimensional tensor. The stress field prediction is a five-dimensional tensor. The preliminary defect distribution is a five-dimensional tensor. .

[0081] in This represents the coordinate position of a voxel in three-dimensional space. For hierarchical index, Let x, y, and z represent the normal stresses in the x, y, and z directions, respectively. These represent the shear stresses in the x, y, and z directions, respectively. This is the index of the stress component, corresponding to one of the six stress components. This is an index for the defect type, corresponding to one of the four types of defects.

[0082] S24. Update the network parameters of the adversarial generative graph attention network based on the preliminary prediction results.

[0083] Step S24 further includes:

[0084] S241. Input the preliminary defect distribution and the real defect distribution samples into the discriminator respectively, and calculate the authenticity score through multi-layer graph convolution. Based on the authenticity score, use the Wasserstein distance metric to predict the difference between the distribution and the real distribution to obtain the adversarial loss value.

[0085] In step S24, the present invention extracts the preliminary defect distribution from the generator output layer and inputs it into the discriminator simultaneously with the labeled real defect distribution samples. For the real defect distribution samples, the CT scan results from the previous standard samples have labeled the real defect type and location of each voxel. After input, the first layer of the discriminator's graph convolution receives the graph representation of the defect distribution, with each node feature being a 4-dimensional defect probability vector; the second to fifth layers of graph convolution extract more abstract discriminative features layer by layer, with the convolution kernel size set to a 3×3×3 voxel neighborhood, and the number of output channels for each layer being 64, 128, 256, and 512 respectively; after the fifth layer output, global average pooling compresses all node features into a 512-dimensional vector, and outputs a authenticity score through two fully connected layers (512→128→1) and a Sigmoid activation function. A score close to 1 indicates that the discriminator considers the input distribution to be from a real sample, while a score close to 0 indicates that it was generated by the generator.

[0086] The Wasserstein distance is then calculated, and the expression is defined as:

[0087]

[0088] in, This represents the Wasserstein distance between the two parameters. For the discriminator output, For the true distribution, To generate a distribution, It expresses expectation.

[0089] Furthermore, the resistance loss value Defined as:

[0090]

[0091] The second term is the gradient penalty term. For random interpolation of the true distribution and the generated distribution, The value is 10, and the gradient penalty forces the discriminator to satisfy the 1-Lipschitz continuity constraint.

[0092] S242. Substitute the predicted temperature field into the Navier-Stokes equation to calculate the temperature residual, and substitute the predicted stress field into the thermoelastic constitutive equation to calculate the stress residual. Perform norm integration on the temperature residual and the stress residual in the spatial and time domains to obtain the physical constraint loss value.

[0093] In step S242, the present invention predicts the temperature field. Substituting into the three-dimensional unsteady heat conduction equation, the expression is:

[0094]

[0095] in For material density, For specific heat capacity, Thermal conductivity, This is the laser thermal source term. In the formula, the time derivative on the left-hand side of the equation is calculated at each voxel location. Using a first-order backward difference approximation, the spatial gradient on the right side... The calculation is performed using the central difference scheme.

[0096] Among them, laser heat source item According to the Gaussian distribution model, the expression is:

[0097]

[0098] in For absorption rate, For laser power, The radius of the light spot is... This represents the distance from the center of the light spot. The numerical differences between the left and right sides of the equation are calculated at all voxel positions to obtain the temperature residual field. .

[0099] Subsequently, the stress field prediction will be... Substituting into the thermoelastic constitutive equation, we get:

[0100]

[0101] in For elastic tensors, For tensor contraction operations, The total strain tensor is calculated from the displacement field gradient. The Johnson-Cook model was used to calculate the plastic strain tensor. The thermal strain tensor is expressed as:

[0102]

[0103] in, The coefficient of thermal expansion is The temperature value at the current moment. For reference temperature, It is a unit tensor.

[0104] Subsequently, this invention substitutes the predicted stress into the right side of the equation at each voxel location to calculate the strain, then calculates the stress back based on the strain, and calculates the difference between the predicted stress and the back-calculated stress to obtain the stress residual field. Subsequently, the temperature residual field and stress residual field were analyzed throughout the entire space domain. and time domain Perform L2 norm integration to calculate the physical constraint loss value. The expression is:

[0105]

[0106] in, For the temperature residual field, For the stress residual field, and These are the balance coefficients, with values ​​of 0.3 and 0.7 respectively.

[0107] S243. The adversarial loss value and the physical constraint loss value are weighted and summed, and a regularization term is added to construct the total loss function.

[0108] Furthermore, in step S243, the present invention uses the adversarial loss value With the physical constraint loss value Perform a weighted summation, with weighting coefficients Set to 0.5 to construct the composite loss. The expression is:

[0109]

[0110] Then, an L2 regularization term was added. The expression is:

[0111]

[0112] in and These represent all trainable parameters for the generator and discriminator, respectively, with δ being the value of the regularization coefficient. .

[0113] The final total loss function is:

[0114]

[0115] in, This is the total loss function.

[0116] S244: Calculate the gradient of the total loss function with respect to the parameters of the generator and the discriminator through the optimizer, update the network parameters of the adversarial generative graph attention network using a learning rate decay strategy, and determine convergence when the rate of change of the total loss function is less than a set threshold.

[0117] Furthermore, in step S244, the present invention uses the Adam optimizer to update the parameters of the generator and discriminator using gradient descent, respectively. Specifically, the total loss function is first calculated. For generator parameters gradient and discriminator parameters gradient The gradient is propagated layer by layer using the backpropagation algorithm. The initial learning rate for the generator is set to 0.0002, and the initial learning rate for the discriminator is set to 0.0001, with a ratio of 2:1 to balance adversarial training. The learning rate decay strategy uses cosine annealing, decreasing to 0.9 times the current value every 50 epochs. The network parameter update formula is:

[0118]

[0119] in, For the first The set of network parameters for each step For the first The set of network parameters for each step For the first The learning rate of the step. and These are the first-order moment estimate and the second-order moment estimate of the gradient, respectively. Take numerical stability terms During training, the total loss function value is recorded every 10 iterations. The relative rate of change of the total loss function over 50-100 consecutive iterations is then recorded. The expression is:

[0120]

[0121] in, Let be the total loss function at time t. The total loss function at time t-1. When the relative rate of change of the total loss function is less than a set threshold... At that point, the network training is considered to have converged.

[0122] S25. The difficulty of the training samples is adjusted by adopting a course learning strategy. The generator outputs a four-dimensional tensor form of the defect probability distribution field, labeling the porosity probability, crack probability, inclusion probability and segregation probability of each voxel.

[0123] Furthermore, this invention employs a course-based learning strategy to organize training data, categorizing standard sample data into three levels of difficulty: easy samples are single-pass, single-layer scans with stable process parameters; medium samples are multi-pass overlapping data with process parameters fluctuating by ±5%; and difficult samples are complex blade profiles with process parameters fluctuating by ±10%. In the early training phase (epochs 1-30), only easy samples are used; in the middle phase (epochs 31-70), easy and medium samples are mixed; and in the later phase (epochs 71-100), difficult samples are added.

[0124] Finally, the converged generator outputs a defect probability distribution field in the form of a four-dimensional tensor. Where x, y, and z are voxel space coordinates, It includes four categories: pores, cracks, inclusions, and segregation, at each voxel location. Output vector , representing the probability of each type of defect occurring in the voxel, respectively. The sum of the four probability values ​​is not necessarily equal to 1 because defect types are not mutually exclusive. The confidence level is also output. ,in Information entropy, denoted as , among which is Entropy is the predicted probability of a certain type of defect. The lower the entropy value, the more certain the prediction and the higher the confidence level.

[0125] S3. Based on the defect probability distribution field, the process parameter vector is optimized through a Bayesian optimization algorithm to obtain an adaptive control parameter combination. The manufacturing process is then controlled in zones according to the adaptive control parameter combination to obtain the zone control result.

[0126] Step S3 further includes:

[0127] S31. Defect suppression degree, forming efficiency, surface roughness and residual stress are set as optimization objectives, and laser power, scanning speed, spot diameter, powder feeding rate, protective gas flow rate, substrate temperature, scanning strategy angle, overlap rate and layer thickness are used as decision variables to construct a multi-objective optimization problem.

[0128] This invention constructs a multi-objective optimization problem based on the aforementioned defect probability distribution field. First, four optimization objective functions are defined: the defect suppression degree objective function... By summing the comprehensive defect probability of each voxel in the defect probability distribution field. With the volume of this voxel The product of the products, divided by the total volume of the blades. Finally, subtract the ratio from 1 to obtain the forming efficiency objective function. By scan speed With powder utilization rate The product divided by the total length of the scan path Compared to single-layer manufacturing time The product is calculated to obtain the powder utilization rate. Based on actual sediment mass Calculated by dividing by the product of powder mass flow rate and scanning time; surface roughness objective function By calculating the actual surface height With nominal surface height The root mean square of the difference is obtained; residual stress objective function. By analyzing the three principal stress components in the predicted stress field , , The equivalent stress is calculated according to the Von Mises criterion, then integrated over the entire blade volume domain and divided by the total volume to obtain the result.

[0129] Subsequently, nine decision variables were defined to form a process parameter vector. Let F1, F2, F3, and F4 represent the laser power, scanning speed, spot diameter, powder feed rate, protective gas flow rate, substrate temperature, scanning strategy angle, overlap ratio, and layer thickness, respectively. The final multi-objective optimization problem is expressed as simultaneously maximizing F1 and F2 and minimizing F3 and F4, with constraints including parameter boundary constraints. Linear energy density constraint Molten pool stability constraints And the matching constraints for sending powder.

[0130] S32. Construct a Gaussian process surrogate model, and encode the quantum dot fluorescence decay time and terahertz absorption coefficient in the multidimensional monitoring feature parameter set into the prior covariance structure of the physical information kernel function to obtain a surrogate model with physical constraints.

[0131] In step S32, the construction of the physical information kernel function includes:

[0132] S321. Define the basic kernel function, which expresses the similarity between two points in the parameter space in the form of radial basis functions, and set the signal variance and multiple feature length scales.

[0133] This invention defines a fundamental kernel function that uses radial basis functions (RBF) to express any two process parameter vectors in the parameter space. and The similarity between them, the basic kernel function form is:

[0134]

[0135] Where d is the dimension index value. The overall amplitude of the output of the signal variance parameter control function. For any process parameter vector In the Specific values ​​in each dimension For another process parameter vector In the Specific values ​​in each dimension The feature length scale parameter for the d-th dimension controls the correlation distance in that dimension. In a specific embodiment, feature length scales are set for the nine dimensions of the process parameter vector: for the laser power dimension... Set to 50W, scanning speed dimension Set to 200mm / s, spot diameter dimension Set to 0.5mm, in the powder feeding rate dimension. Set to 2g / min, protecting the gas flow rate dimension Set to 5L / min, substrate temperature dimension Set to 50℃, scanning strategy angle dimension Set to 30 degrees, overlap rate dimension Set to 0.1, layer thickness dimension Set to 0.02mm. Signal variance. Set to 1.0. Due to the decay property of the exponential function, when the difference between two parameter vectors in a certain dimension exceeds 2-3 times the feature length scale of that dimension, the contribution of that dimension to the kernel function value rapidly approaches zero.

[0136] S322. Extract the quantum dot fluorescence decay time and terahertz absorption coefficient from the set of multidimensional monitoring feature parameters, and construct a physical correction factor. The physical correction factor reflects the physical response characteristics when the process parameter vector deviates from the standard value.

[0137] Furthermore, this invention extracts the quantum dot fluorescence decay time from the 150 feature dimensions of the multidimensional monitoring feature parameter set. and terahertz absorption coefficient Two key physical parameters, one of which is the quantum dot fluorescence decay time, were extracted from the time-domain signal of the bimodal fluorescence spectrum through exponential fitting, under normal process conditions. Reference value Approximately 60 ns; the terahertz absorption coefficient is obtained by calculating the transmittance from the terahertz transmission spectrum using the Beer-Lambert law and then taking the logarithm. Under normal process conditions... Reference value Approximately 15cm -1 Then, a physical correction factor was constructed. This factor quantifies the process parameter vector in a double-exponential form. The physical response characteristics when deviating from the standard value are expressed as:

[0138]

[0139] in and The weighting coefficients were set to 0.5 and 0.3, respectively. When the process parameters caused the fluorescence decay time or terahertz absorption coefficient to deviate from the reference value, the physical correction factor value decreased, reflecting the degree of abnormality in the physical state of the material under this parameter combination. The aim was to establish a system based on the monitoring data of standard samples under different process parameters. and With process parameter vector The mapping relationship.

[0140] S323. Multiply the physical correction factor by the basic kernel function to obtain the physical information kernel function, which is used to construct the covariance matrix of the surrogate model of the Gaussian process, thereby obtaining a surrogate model with physical constraints.

[0141] The present invention will use the physical correction factor Acting on two points in the parameter space respectively and Multiply with the basic kernel function By combining these, we obtain the physical information kernel function. The kernel function constructs a covariance matrix K in the parameter space, where each element represents the first covariance matrix. The sampling point and the first Covariance of each sampling point.

[0142] When there are already n sampling points, the covariance matrix K has dimensions n×n. For new parameter points to be evaluated... Calculate its covariance vector with all sampled points. Subsequently, the covariance matrix K is decomposed into the product of a lower triangular matrix L and its transpose using Cholesky decomposition. ,in This represents the matrix transpose, followed by solving the system of linear equations through forward and backward substitutions. Obtain the weight vector ,in The objective function observations are for the already sampled points. The predicted mean for the new points is calculated as follows: The prediction variance is calculated as follows: Finally, a Gaussian process surrogate model with physical constraints is obtained through the above calculations.

[0143] S33. Based on the multi-objective optimization problem and the surrogate model, the candidate parameter set is used to approximate the posterior distribution of the parameter space through variational inference. The Pareto optimal solution set is output through knowledge gradient balancing sampling and multiple virtual evaluations, and the adaptive control parameter combination is obtained by screening.

[0144] Step S33 further includes:

[0145] S331. Based on the multi-objective optimization problem and relying on the surrogate model, determine the range of candidate parameter combinations in the parameter space and obtain the candidate parameter set.

[0146] In step S331, the present invention is based on the parameter boundary constraints defined in the multi-objective optimization problem. Subsequently, Latin hypercube sampling was used in the nine-dimensional process parameter space to generate initial candidate parameter combinations.

[0147] Specifically, this invention defines the value range for each parameter dimension. The sample is divided into M equal intervals, and Latin hypercube sampling is used to ensure that the sampling points are uniformly distributed and non-clustered in each dimension. Simultaneously, this invention, combined with the defect probability distribution field, identifies the spatial regions corresponding to voxel clusters with high defect probabilities. The sampling density is increased in the parameter subspace near these high-risk regions, generating an additional 20-30 candidate parameter combinations. All candidate parameter combinations are then aggregated to form a candidate parameter set. The set size M is 80-150 parameter combinations. Based on the physically constrained surrogate model, each parameter combination in the candidate parameter set is... A preliminary assessment was conducted, and the predicted mean values ​​for the four objective functions were calculated.

[0148] S332. The variational inference method is used to approximate the posterior distribution of the candidate parameter set, and the expected defect rate mean and 95% confidence interval of multiple candidate parameter combinations are calculated through a surrogate model.

[0149] This invention employs a variational inference method to approximate the posterior distribution of the candidate parameter set. Variational inference introduces a variational distribution... Approximate true posterior distribution Variational distribution Selecting a multivariate Gaussian distribution Subsequently, the KL divergence between the variational distribution and the true posterior distribution is minimized. Optimizing the variational parameters is equivalent to maximizing the lower bound of evidence. The Stochastic Gradient Variational Bayesian (SGVB) algorithm is used to sample parameters from the variational distribution. The surrogate model is used to calculate the predicted value of the objective function under this parameter, and the prediction likelihood is calculated. Accumulated gradient and After 200-500 iterations to update the variational parameters, the converged variational distribution is obtained. It provides an approximation of the posterior distribution of the parameter space. Subsequently, for each parameter combination in the candidate parameter set... The expected defect rate is calculated using a proxy model, where the defect rate is defined as follows: At the same time, calculate the 95% confidence interval of the predicted mean. .

[0150] Of the parameters mentioned above, For observed data, It is the mean vector. Let covariance matrix be the variance matrix. Indicates the expectation. The log-likelihood of observed data measures the parameters. Observational data Explanatory power This indicates the calculation of the KL divergence. The predicted mean of the defect rate. This represents the predicted standard deviation.

[0151] S333. Explore and utilize the parameter space of the function trade-off through knowledge gradient, and prioritize sampling the candidate parameter combination with the largest gain.

[0152] This invention strikes a balance between exploring and utilizing the parameter space through a knowledge gradient acquisition function. (Knowledge gradient acquisition function) Defined as sampling a candidate parameter combination The expected improvement after reaching the global optimum. First, calculate the optimal value of the currently sampled parameter combinations. ,in For the first The predicted mean of each sampled point. Then, for the candidate parameter combinations to be evaluated... Calculate the expected optimal value after sampling this point. The expected value is approximated by Monte Carlo sampling: 500-1000 random numbers are sampled from a standard normal distribution N(0,1). Calculate the conditional optimal value for each sample. The average of all sampling results is obtained. Next, the knowledge gradient value is calculated and represented as... .

[0153] Then, iterate through all parameter combinations in the candidate parameter set, calculate the knowledge gradient value for each combination, and select the parameter combination with the largest KG value. Perform the next virtual evaluation. When there are more than 8 consecutive voxels with crack probabilities in the defect probability distribution field... When the value is greater than 0.6, the parameter subspace corresponding to the voxel cluster is identified. The candidate parameter combinations within the subspace are assigned a sampling weight of 2-3 times, and the candidate parameter combinations with the largest gain are sampled first.

[0154] S334. After multiple virtual evaluations, the candidate parameter combinations and prediction results are iteratively updated, and the Pareto optimal solution set that satisfies the multi-objective optimization constraints is output.

[0155] In step S334, the present invention performs 100-200 virtual evaluation iterations. In each iteration, the parameter combination selected in step S333 is used. The surrogate model is used to calculate its predicted values ​​on the four objective functions. , , , And its prediction uncertainty. Subsequently, this parameter combination and its prediction results are added to the sampled dataset, the training data of the surrogate model is updated, and the covariance matrix is ​​recalculated. and weight vector The updated surrogate model recalculates the predicted mean and variance for the remaining unevaluated parameter combinations in the candidate parameter set, and updates the knowledge gradient values.

[0156] When the knowledge gradient improvement is achieved after 20 consecutive samplings The optimization process is considered convergent when the value is less than 1% of the initial knowledge gradient. From all evaluated parameter combinations, solutions satisfying the multi-objective optimization constraints are selected to construct the Pareto front: parameter combination... Dominate If and only if Not inferior to in all objectives And strictly superior to at least one objective All non-dominated solutions are identified using a non-dominated sorting algorithm, forming a Pareto optimal solution set. The solution set contains 20-50 combinations of parameters, each representing a different objective trade-off.

[0157] S335. Based on the aforementioned defect probability distribution field characteristics, select adaptive control parameter combinations from the Pareto optimal solution set that are suitable for defect conditions in different sub-regions.

[0158] Furthermore, in step S335, the present invention combines the spatial distribution characteristics of the defect probability distribution field to obtain the optimal Pareto solution set. The middle part selects suitable adaptive control parameter combinations for different sub-regions. Specifically, firstly, spatial clustering analysis is performed on the defect probability distribution field. The DBSCAN algorithm is used to identify voxel clusters with defect probabilities higher than a threshold (set to 0.3 in the specific implementation), and voxels with similar defect types and spatial proximity are grouped into the same cluster. For the stomatal tendency region ( (Voxel clusters), parameter combinations with lower laser power and slower scanning speed are selected from the Pareto optimal solution set. These parameters reduce the degree of molten pool overheating; targeting crack-prone regions ( (voxel clusters), screening for parameter combinations with higher substrate temperatures and larger scanning strategy angle changes, these parameters reduce temperature gradients and residual stress; targeting areas prone to non-fusion ( (voxel clusters), screening for parameter combinations with higher laser power and lower powder feed rate, these parameters increase energy density; targeting segregation-prone regions ( (Voxel clusters) are used to screen for parameter combinations with faster scanning speeds and lower overlap rates, which shorten the melt pool duration. For normal regions with low defect probabilities (overall defect probability < 0.2), the forming efficiency is screened. A higher parameter combination. Through the above matching rules, this invention ultimately assigns 1-3 candidate adaptive control parameter combinations to each spatial region.

[0159] S34. Based on the defect probability distribution field, the scanning path is divided into several sub-regions. For sub-regions where the defect probability exceeds the threshold, corresponding adaptive control parameter combinations are assigned to perform partitioned asynchronous control to obtain partitioned control results.

[0160] In step S34, the present invention divides the scanning path into several sub-regions based on the defect probability distribution field. First, the scanning path is spatially divided into a 5mm × 5mm grid, with each grid cell representing a potential sub-region. The average defect probability of all voxels within each grid cell is then calculated. ,when When the value exceeds the threshold of 0.25, the grid cell is marked as a high-risk sub-region; when Regions with a risk level below 0.15 are marked as low-risk sub-regions; those with a risk level between 0.15 and 0.15 are marked as medium-risk sub-regions.

[0161] For high-risk sub-regions, the defect suppression adaptive control parameter combination selected in step S335 is assigned. This combination... It scores highly on the target; for low-risk sub-regions, it assigns an efficiency-priority parameter combination, which in... For high-risk sub-regions, a balanced parameter combination is assigned. A 3-5 layer (approximately 0.1-0.15 mm) gradient transition zone is set between adjacent sub-regions. The process parameters within the transition zone smoothly transition from the parameter combination of one sub-region to the parameter combination of the adjacent sub-region using a linear interpolation method, avoiding process instability caused by abrupt parameter changes.

[0162] In the actual manufacturing process, when the laser scans a certain sub-region, the control system automatically calls the corresponding adaptive control parameter combination for that sub-region to adjust parameters such as laser power and scanning speed in real time. During further asynchronous regional control, the multi-physics sensing system continuously collects five-dimensional raw monitoring data streams to verify the parameter adjustment effect and obtain the regional control results.

[0163] S4. For the residual defects in the partition control results, a reinforcement learning strategy network is used to select a repair action sequence. The repair action sequence is input into the execution unit to perform cross-level collaborative processing on the residual defects to obtain the repaired blade structure.

[0164] Step S4 further includes:

[0165] S41. Construct a partially observable Markov decision process model, defining the multidimensional monitoring feature parameter set of the current layer and the multilayers below it as the state space, defining local remelting, laser shock strengthening, ultrasonic micro-forging, arc-enhanced deposition and cold spray filling and their sub-parameters as the action space, and defining the weighted combination of defect elimination rate and penalty for introducing new defects as the reward function.

[0166] Furthermore, this invention constructs a partially observable Markov decision process (POMDP) ​​model to address the residual defects still present in the zoning control results. The state space is defined as the set of multi-dimensional monitoring feature parameters of the current manufacturing layer and the 3-8 layers already formed below it. Specifically, it includes 150-dimensional feature vectors of all voxels in the current layer and the 7 layers below it (a total of 8 layers). Each layer has approximately 3 million voxels, resulting in a state space dimension of 8 × 3 million × 150 for 8 layers. To reduce dimensionality, principal component analysis (PCA) is used to compress the 150-dimensional features into 30 key features. The state space is represented as follows: ,in For the first The compressed feature matrix of the layer.

[0167] The action space is defined as five types of repair strategies and their sub-parameters, including local remelting (parameters include remelting power, remelting speed, and number of remelting cycles), laser shock peening (parameters include shock energy density, pulse width, and number of shocks), ultrasonic micro-forging (parameters include vibration frequency, amplitude, and pressure), arc-enhanced deposition (parameters include current, wire feed speed, and protective gas composition), and cold spray filling (parameters include powder particle size, spraying speed, and spraying temperature). Each repair strategy contains 8-12 adjustable sub-parameters, resulting in a total action space of approximately 50 dimensions, represented as follows: ,in For repair strategy types, This is the corresponding sub-parameter vector.

[0168] The reward function is defined as follows: ,in The defect elimination rate is calculated by the ratio of defect volume before and after repair. The probability of introducing new defects is assessed by incrementally estimating the probability of defects around the repair area. and The weighting coefficients were set to 0.7 and 0.3 respectively.

[0169] S42. Establish a deep deterministic policy gradient network, which includes a two-stream Actor network and a Critic network. Input the state space into the convolutional stream and recurrent stream of the two-stream Actor network to process spatial features and temporal dependencies respectively, and output continuous repair parameters. The Q value of the state-action pair is evaluated by the Critic network.

[0170] This invention establishes a Deep Deterministic Policy Gradient (DDPG) network to process repair decisions in a continuous action space. The DDPG network comprises four sub-networks: a two-stream Actor network, a target Actor network, a Critic network, and a target Critic network.

[0171] The dual-stream Actor network employs a dual-channel input architecture. The convolutional stream receives spatial features from the state space S_t, with an input dimension of 3 million × 30 (voxels × feature dimension). It extracts spatial patterns through three layers of 3D convolution (3×3×3 kernels, stride 2, number of channels 64→128→256) and outputs a 256-dimensional spatial feature vector. The recurrent stream receives the temporal features of the state space, inputs the data from the 8 layers into the GRU unit (128 hidden layer dimensions) in chronological order, captures the cross-layer evolution trend, and outputs a 128-dimensional temporal feature vector. After obtaining two feature vectors, this invention concatenates the two feature vectors and fuses them through a fully connected layer (384→256→128). The output layer contains two branches: a repair strategy type branch selects a 5-dimensional probability distribution through softmax output. Repair parameter branch outputs 50-dimensional continuous repair parameters After tanh activation, it is mapped to the interval [-1,1] and then linearly transformed to the actual parameter range.

[0172] Critic network receive state-action pair The state is encoded using a three-layer fully connected layer (384→256→128), and the action is encoded using a two-layer fully connected layer (50→64→128). The concatenation of the two encodings is then processed through a two-layer fully connected layer (256→128→1) to output the Q-value. Evaluate the long-term cumulative reward of taking this action in this state.

[0173] The target Actor network and target Critic network have the same structure as their corresponding main networks, and their parameters are slowly copied from the main network through soft updates.

[0174] S43. Load the pre-trained multi-material repair strategy library through the meta-learning module, perform rapid adaptation with few samples based on the multi-dimensional monitoring feature parameter set of the current blade preceding layer, and output the repair action sequence for the residual defect by the dual-stream Actor network.

[0175] In step S43, the present invention achieves rapid adaptation across materials and defect types through a meta-learning module. This meta-learning module stores a pre-trained multi-material repair strategy library, containing strategy parameters for 100-150 material-defect-process combinations. Each combination corresponds to a set of trained DDPG network parameters. The strategy library structure is {(nickel-based superalloy-porosity-laser cladding, W1), (nickel-based superalloy-crack-laser cladding, W2), (titanium alloy-porosity-electron beam melting, W3),...}.

[0176] For a new turbine blade manufacturing task, extract metadata such as material composition, defect type distribution, and manufacturing process of the current blade. Retrieve the 5-10 tasks with the highest similarity in the strategy library and calculate the similarity index. ,in Material composition similarity is calculated using the cosine distance between chemical composition vectors. The similarity of defect distributions is calculated using the KL divergence of the defect probability distribution field. The similarity of process parameters is calculated using the Euclidean distance between parameter vectors. , , These are the corresponding weight values.

[0177] Subsequently, this invention uses the network parameters of the task with the highest similarity as initial weights and fine-tunes them based on the actual manufacturing data of the first 5-15 layers of the current blade. The fine-tuning process uses gradient descent to update only the parameters of the last 3-5 layers of the dual-stream Actor network and Critic network, while freezing the parameters of the previous layers to retain pre-trained knowledge.

[0178] For each layer, 10-20 state-action-reward samples are generated and stored in the experience replay buffer. Samples of size 32 are randomly sampled from the buffer for training, and the Actor loss is calculated. and Critic loss The parameters are updated through backpropagation. Of the parameters mentioned above, Indicates the expectation. The current state of the environment. The action taken by the agent in this state. The reward obtained after performing the action. The new state transitioned to after performing an action. Indicates the state of the Critic network. The action that executes the Actor's output. Value assessment For the target Critic network to adapt to the new state The action output by the next target Actor Value assessment, For the current state of the Critic network Next action Value prediction.

[0179] Fast adaptation is completed when the training loss converges to less than 10% of the initial value. The adapted two-stream Actor network takes the current layer state space as input and outputs a sequence of repair actions for the residual defects. The sequence length T is determined based on the number and complexity of defects, and consists of 3-8 consecutive actions.

[0180] S44. Drive the femtosecond laser probe to scan along the path of the residual defect to obtain the defect depth and direction data. Perform multi-stage combined processing on the defect area corresponding to the residual defect according to the repair strategy determined by the repair action sequence to obtain the repaired blade structure.

[0181] This invention uses a femtosecond laser probe to accurately characterize the residual defects. The femtosecond laser probe emits ultrashort laser pulses with a pulse width of 100 fs and a repetition frequency of 80 MHz, scanning along the predicted path of the residual defects. Using two-photon fluorescence imaging, when the femtosecond laser is focused on the defect location, stress concentration at the defect boundary leads to a 2-3 fold increase in two-photon fluorescence intensity. The three-dimensional contour of the defect is reconstructed by point-by-point scanning. Measurement data includes the defect depth. (Dimensions along the z-direction), Defect orientation (Angle between the defect axis and the scanning direction), defect length and width .

[0182] Subsequently, according to the repair action sequence The first action The repair strategy type and parameters were determined. Specifically, for crack defects deeper than 0.5 mm, a four-stage combined repair strategy was adopted: The first stage, preheating, involved using a 200W continuous laser to scan a 5 mm radius around the defect for 10 seconds, raising the local temperature to 200℃; the second stage, ultrasonic micro-forging, involved applying vibrations at a frequency of 40 kHz, an amplitude of 15 μm, and a pressure of 50 MPa to the defect surface using an ultrasonic transducer for 5 seconds, causing the crack surface to close and generating compressive stress; the third stage, filling and remelting, involved switching to pulsed laser mode with a peak power of 800 W, a pulse width of 1 ms, and a duty cycle of 0.3, scanning along the crack path for filling and melting, with the powder feed rate reduced to 60% of the normal value to control the dilution rate; the fourth stage, laser shock peening, involved applying an energy density of 10 GW / cm² to the remelting area. 2 A laser shock with a pulse width of 8 ns is applied three times, with a 2-second interval between each shock, inducing a residual compressive stress field on the material surface to counteract tensile stress. For pore defects with a diameter less than 0.3 mm, a local remelting strategy is adopted, with the power reduced to 50% of the normal value, the speed increased to 3 times, and the remelting depth controlled at 0.05-0.1 mm. During the repair process, the invention continuously monitors the process using the aforementioned multiphysics sensing system. After the repair is completed, the density is verified by terahertz transmission detection, and the porosity is calculated. If the porosity exceeds 0.3%, the next action is called from the repair action sequence. If the repair parameters are adjusted and the process is repeated until the porosity is satisfactory or the action sequence is exhausted, the repaired blade structure is finally obtained.

[0183] like Figure 2 As shown, the present invention also provides a defect correction system for additive manufacturing of aerospace engine turbine blades, comprising:

[0184] Acquisition module 100: used to collect signal data during the additive manufacturing process of quantum dot labeling agent adsorbed on the surface of metal powder, and obtain a set of multi-dimensional monitoring feature parameters;

[0185] Prediction module 200: Used to receive a set of multi-dimensional monitoring feature parameters as node feature input, and to predict the spatiotemporal evolution of the molten pool region through an adversarial training mechanism to obtain the defect probability distribution field;

[0186] Optimization module 300: used to perform multi-objective optimization of the process parameter vector based on the defect probability distribution field using a Bayesian optimization algorithm to obtain an adaptive control parameter combination;

[0187] Execution module 400: is used to perform partitioned control of the manufacturing process according to the adaptive control parameter combination to obtain partitioned control results; it is also used to execute the repair action sequence output by the repair module to perform cross-level collaborative processing of residual defects;

[0188] Repair module 500: Used to select a repair action sequence using a reinforcement learning strategy network to address the residual defects in the partition control results.

[0189] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0190] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the defect correction method for additive manufacturing of aerospace engine turbine blades described in various embodiments or some parts of embodiments.

[0191] This invention achieves advanced perception of nanoscale defects through the fusion of quantum dot labeling and multiphysics sensing, improving the spatiotemporal resolution of defect detection. Secondly, it uses an adversarial generative graph attention network combined with physical constraints to accurately predict the evolution of defects at levels 3-10, achieving extremely high prediction accuracy and transforming defect intervention from passive response to proactive prevention. Furthermore, this invention integrates a physical information kernel function with a variational Bayesian optimization algorithm for adaptive process control, reducing defect incidence, improving surface roughness, and lowering residual stress. Additionally, this invention uses a reinforcement learning policy network combined with a meta-learning module for intelligent repair of residual defects, improving repair success rate, increasing the yield of turbine blade additive manufacturing, shortening the manufacturing cycle, and improving material utilization. This provides a technical solution for high-quality intelligent manufacturing of key components for aerospace engines.

[0192] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for defect correction in additive manufacturing of aerospace engine turbine blades, characterized in that, include: S1. Adsorb quantum dot labeling agents onto the surface of metal powder, and collect signal data during the additive manufacturing process through a multi-physics field sensor to obtain a set of multi-dimensional monitoring characteristic parameters. S2. Construct a graph neural network model, take the multi-dimensional monitoring feature parameter set as node feature input, and predict the spatiotemporal evolution of the molten pool region through an adversarial training mechanism to obtain the defect probability distribution field. Step S2 further includes: S21, discretizing the blade manufacturing space into a voxel grid, using each voxel as a graph node, and establishing spatial adjacency edges and physical causal edges between nodes based on the heat conduction direction, stress propagation path, and melt flow trajectory to obtain a spatiotemporal dynamic graph structure; S22, constructing an adversarial generative graph attention network, which includes a generator, a discriminator, and a physical constraint module; S23, assigning the multidimensional monitoring feature parameter set to the nodes of the spatiotemporal dynamic graph structure, inputting it into the generator for multi-head attention aggregation calculation, and obtaining preliminary prediction results; the preliminary prediction results include: temperature field prediction, stress field prediction, and preliminary defect distribution; S24, updating the network parameters of the adversarial generative graph attention network based on the preliminary prediction results; S25, adjusting the difficulty of training samples using a course learning strategy, and having the generator output a four-dimensional tensor form of defect probability distribution field, labeling the porosity probability, crack probability, inclusion probability, and segregation probability of each voxel; S3. Based on the defect probability distribution field, the process parameter vector is optimized through a Bayesian optimization algorithm to obtain an adaptive control parameter combination. The manufacturing process is then controlled in zones according to the adaptive control parameter combination to obtain the zone control result. S4. For the residual defects in the partition control results, a reinforcement learning strategy network is used to select a repair action sequence. The repair action sequence is input into the execution unit to perform cross-level collaborative processing on the residual defects to obtain the repaired blade structure. Step S4 further includes: S41, constructing a partially observable Markov decision process model, defining the multidimensional monitoring feature parameter set of the current layer and the multi-layers below it as the state space, defining local remelting, laser shock strengthening, ultrasonic micro-forging, arc-enhanced deposition, and cold spray filling and their sub-parameters as the action space, and defining the weighted combination of defect elimination rate and penalty for introducing new defects as the reward function; S42, establishing a deep deterministic policy gradient network, the deep deterministic policy gradient network including a two-stream Actor network and a Critic network, and inputting the state space into the convolutional flow and recurrent loop of the two-stream Actor network. The streams process spatial features and temporal dependencies respectively, outputting continuous repair parameters. The Q-value of the state-action pair is evaluated by the Critic network. S43: A pre-trained multi-material repair strategy library is loaded through the meta-learning module. Based on the multi-dimensional monitoring feature parameter set of the current blade preceding layer, a few-sample fast adaptation is performed. The repair action sequence for the residual defect is output by the dual-stream Actor network. S44: The femtosecond laser probe is driven to scan along the path of the residual defect to obtain defect depth and orientation data. According to the repair strategy determined by the repair action sequence, the defect region corresponding to the residual defect is processed in a multi-stage combination to obtain the repaired blade structure.

2. The method of claim 1, wherein: Step S1 further includes: S11. Core-shell quantum dots are adsorbed onto the surface of nickel-based superalloy powder by electrostatic self-assembly to obtain functionalized labeled powder. S12. Collect the bimodal fluorescence spectrum, terahertz transmission spectrum, eddy current conductivity signal, laser ultrasonic echo signal and magneto-optical Kerr signal of the functionalized labeled powder under laser irradiation to obtain a five-dimensional raw monitoring data stream. S13. Perform Fourier transform and wavelet multi-scale decomposition on the five-dimensional original monitoring data stream to extract frequency domain energy distribution, phase delay, time-frequency localization features and singular values ​​to obtain a multi-dimensional monitoring feature parameter set.

3. The method of claim 2, wherein: In step S11, the core-shell structured quantum dot is a CdSe / ZnS or InP / ZnS material system.

4. The method of claim 1, wherein: Step S24 further includes: S241. Input the preliminary defect distribution and the real defect distribution samples into the discriminator respectively, and calculate the authenticity score through multi-layer graph convolution. Based on the authenticity score, use the Wasserstein distance metric to predict the difference between the distribution and the real distribution to obtain the adversarial loss value. S242. Substitute the temperature field prediction into the Navier-Stokes equation to calculate the temperature residual, and substitute the stress field prediction into the thermoelastic constitutive equation to calculate the stress residual. Perform norm integration on the temperature residual and the stress residual in the spatial domain and the time domain to obtain the physical constraint loss value. S243. The adversarial loss value and the physical constraint loss value are weighted and summed, and a regularization term is added to construct the total loss function; S244: Calculate the gradient of the total loss function with respect to the parameters of the generator and the discriminator through the optimizer, update the network parameters of the adversarial generative graph attention network using a learning rate decay strategy, and determine convergence when the rate of change of the total loss function is less than a set threshold.

5. A defect correction method for additive manufacturing of aerospace engine turbine blades according to claim 1, characterized in that, Step S3 further includes: S31. Set the defect suppression degree, forming efficiency, surface roughness and residual stress as optimization objectives, and use laser power, scanning speed, spot diameter, powder feeding rate, protective gas flow rate, substrate temperature, scanning strategy angle, overlap rate and layer thickness as decision variables to construct a multi-objective optimization problem. S32. Construct a Gaussian process surrogate model, and encode the quantum dot fluorescence decay time and terahertz absorption coefficient in the multidimensional monitoring feature parameter set into the prior covariance structure of the physical information kernel function to obtain a surrogate model with physical constraints. S33. Based on the multi-objective optimization problem and the surrogate model, the candidate parameter set is used to approximate the posterior distribution of the parameter space through variational inference. The Pareto optimal solution set is output through knowledge gradient balancing sampling and multiple virtual evaluations, and the adaptive control parameter combination is obtained by screening. S34. Based on the defect probability distribution field, the scanning path is divided into several sub-regions. For sub-regions where the defect probability exceeds the threshold, corresponding adaptive control parameter combinations are assigned to perform partitioned asynchronous control to obtain partitioned control results.

6. The method of claim 5, wherein: Step S32 further includes: S321. Define the basic kernel function, express the similarity between two points in the parameter space in the form of radial basis functions, and set the signal variance and multiple feature length scales. S322. Extract the quantum dot fluorescence decay time and terahertz absorption coefficient from the set of multidimensional monitoring feature parameters, and construct a physical correction factor. The physical correction factor reflects the physical response characteristics when the process parameter vector deviates from the standard value. S323. Multiply the physical correction factor by the basic kernel function to obtain the physical information kernel function, which is used to construct the covariance matrix of the surrogate model of the Gaussian process, thereby obtaining a surrogate model with physical constraints.

7. The method of claim 5, wherein: Step S33 further includes: S331. Based on the multi-objective optimization problem and relying on the surrogate model, determine the range of candidate parameter combinations in the parameter space and obtain the candidate parameter set; S332. The variational inference method is used to approximate the posterior distribution of the candidate parameter set, and the expected defect rate mean and 95% confidence interval of multiple candidate parameter combinations are calculated through a surrogate model. S333. Explore and utilize the parameter space of function trade-offs through knowledge gradients, and prioritize sampling candidate parameter combinations with the largest gain. S334. After multiple virtual evaluations, the candidate parameter combinations and prediction results are iteratively updated, and the Pareto optimal solution set that satisfies the multi-objective optimization constraints is output. S335. Based on the aforementioned defect probability distribution field characteristics, select adaptive control parameter combinations from the Pareto optimal solution set that are suitable for defect conditions in different sub-regions.

8. An aerospace engine turbine blade additive manufacturing defect correction system for performing an aerospace engine turbine blade additive manufacturing defect correction method according to any one of claims 1 to 7, characterized in that, include: Acquisition module: used to collect signal data during the additive manufacturing process of quantum dot labeling agents adsorbed on the surface of metal powder, and obtain a set of multi-dimensional monitoring characteristic parameters; Prediction module: It receives a set of multi-dimensional monitoring feature parameters as node feature inputs, and performs spatiotemporal evolution prediction of the molten pool region through an adversarial training mechanism to obtain the defect probability distribution field; The prediction module is further used to: discretize the blade manufacturing space into a voxel grid, using each voxel as a graph node, and establish spatial adjacency edges and physical causal edges between nodes based on the heat conduction direction, stress propagation path, and melt flow trajectory to obtain a spatiotemporal dynamic graph structure; construct an adversarial generative graph attention network, which includes a generator, a discriminator, and a physical constraint module; assign the multidimensional monitoring feature parameter set to the nodes of the spatiotemporal dynamic graph structure, input it into the generator for multi-head attention aggregation calculation, and obtain preliminary prediction results; The preliminary prediction results include: temperature field prediction, stress field prediction, and preliminary defect distribution; the network parameters of the adversarial generative graph attention network are updated based on the preliminary prediction results; the difficulty of the training samples is adjusted using a course learning strategy, and the generator outputs a four-dimensional tensor form of defect probability distribution field, labeling the porosity probability, crack probability, inclusion probability, and segregation probability of each voxel; Optimization module: used to perform multi-objective optimization of the process parameter vector based on the defect probability distribution field using a Bayesian optimization algorithm to obtain an adaptive control parameter combination; Execution module: used to perform zoned control of the manufacturing process according to the adaptive control parameter combination to obtain zoned control results; also used to execute the repair action sequence output by the repair module to perform cross-level collaborative processing of residual defects; Repair module: Used to select repair action sequences using a reinforcement learning strategy network to address residual defects in the partitioning control results; The repair module is further used to: construct a partially observable Markov decision process model, defining the multidimensional monitoring feature parameter set of the current layer and the multilayer layers already formed below as the state space, defining local remelting, laser shock strengthening, ultrasonic micro-forging, arc-enhanced deposition, and cold spray filling and their sub-parameters as the action space, and defining the weighted combination of defect elimination rate and penalty for introducing new defects as the reward function; and establish a deep deterministic policy gradient network, which includes a two-stream Actor network and a Critic network, inputting the state space into the convolutional and cyclic flows of the two-stream Actor network. Spatial features and temporal dependencies are processed separately to output continuous repair parameters. The Q-value of the state-action pair is evaluated by the Critic network. A pre-trained multi-material repair strategy library is loaded through the meta-learning module. Based on the multi-dimensional monitoring feature parameter set of the current blade's preceding layer, rapid adaptation with few samples is performed. The repair action sequence for the residual defect is output by the dual-stream Actor network. A femtosecond laser probe is driven to scan along the path of the residual defect to obtain defect depth and orientation data. According to the repair strategy determined by the repair action sequence, the defect region corresponding to the residual defect is processed in multiple stages to obtain the repaired blade structure.

Images