Mechanism-data fusion driven real-time quality control method for additive repair

Abstract

This invention discloses a real-time quality control method for additive repair based on mechanism-data fusion. First, the measured and simulated temperature fields of the repair area are acquired, and a heat flux matrix is ​​constructed by combining laser path planning parameters to form measured and simulated datasets. Then, a dual-branch ConvLSTM encoder is constructed to extract deep spatiotemporal features of the temperature field and heat flux field respectively. The dynamic guidance response of heat flux to temperature evolution is calculated through a multi-head cross-attention mechanism, and adaptive fusion is performed using learnable residual gating parameters. The predicted temperature field for the next moment is reconstructed by the decoder. A two-stage strategy of "simulation pre-training - measured fine-tuning" is adopted for the molten pool temperature field prediction model. Finally, the real-time acquired temperature field is input into the molten pool temperature field prediction model to invert the molten pool morphology, calculate the volume and dilution rate deviation, and output laser power compensation commands in real time through a fuzzy controller, feeding them to the directional energy deposition laser to achieve closed-loop quality control of the repair process.

Description

Technical Field

[0001] This invention belongs to the field of directional energy deposition process monitoring and closed-loop control, specifically a real-time quality control method for additive repair based on mechanism-data fusion. Background Technology

[0002] Directed energy deposition (DED) technology, with its advantages of high deposition efficiency, small size limitations, and gradient manufacturing capability, has shown great potential in the surface strengthening and damage repair of large and complex aerospace components. This repair process involves rapid melting and solidification under extreme temperature gradients, and the repaired area exhibits strong transient and non-stationary characteristics due to its geometrically irregular shape and significant heat accumulation. Particularly in the additive repair of large load-bearing components such as aircraft landing gear and large thin-walled structural parts, the impact of these transient thermal behaviors on microstructure and geometric integrity is particularly significant. Real-time prediction of the temperature field in the repaired area has become a key means of evaluating the quality of additive repair. Studies have shown that the dynamic evolution of the molten pool temperature field is closely related to the dilution rate of the molten pool elements, internal defects in the repaired part, and forming accuracy. Real-time monitoring methods based on multi-source sensing information have shown initial feasibility in additive manufacturing quality assessment, but most methods are still in the theoretical research stage and cannot directly meet the needs of high-standard aerospace industrial production.

[0003] Current mainstream monitoring methods mostly rely on purely data-driven deep learning models. While these models possess strong nonlinear fitting capabilities, they often neglect physical mechanisms such as energy conservation in heat conduction and the dynamic changes of material properties with temperature. This results in weak generalization ability under complex operating conditions, and the prediction results may violate fundamental thermodynamic laws. Furthermore, existing models are often limited to the static characteristics of the molten pool at a single moment, or only calculate statistical indicators such as molten pool area and average temperature. They struggle to accurately characterize the instantaneous dynamic behavior of the molten pool under high-frequency laser heat input, and also fail to quantify the dynamic coupling effect of heat input, heat diffusion, and temperature field over long periods. At the process parameter optimization level, even with the introduction of reinforcement learning methods, the heat accumulation effect is often ignored, leading to poor parameter adaptability in multi-stage repairs. Because real-time closed-loop feedback and compensation from the predicted temperature field to process parameters cannot be achieved, the repair process is highly susceptible to quality defects due to volume deviations or abnormal dilution rates.

[0004] In summary, based on existing technologies, although some studies have attempted to introduce temperature field prediction or process monitoring methods, there is still a general lack of a systematic approach that combines physical mechanisms, real-time prediction, and closed-loop control of process parameters. Summary of the Invention

[0005] To address the above technical problems, this invention proposes a real-time quality control method for additive repair based on mechanism-data fusion, aiming to achieve high-precision real-time prediction and active control of the molten pool morphology and quality during the repair process. This solves the problem of difficulty in precise control due to the complexity of dynamic thermal processes in laser additive repair, and ultimately achieves intelligent closed-loop quality control of the repair process.

[0006] A mechanism-data fusion-driven real-time quality control method for additive repair includes the following steps: Step 1: Obtain the measured temperature field matrix of the repair area through infrared thermal imaging, and generate the corresponding heat flux matrix by combining the laser path planning parameters. Align the temperature field matrix and the heat flux matrix in terms of spatiotemporal features to construct a training dataset. Step 2: Input the temperature field matrix from the historical time to the current time and the heat flux matrix for the next time into the pre-trained molten pool temperature field prediction model to obtain the temperature field for the next time. The molten pool temperature field prediction model includes: a dual-branch feature encoder module, a multi-head cross-attention physical guidance module, a feature fusion module, and a decoder; The dual-branch feature encoder module consists of a temperature field encoder. With thermal flux encoder The process involves performing ConvLSTM temporal recurrent convolution on the temperature field matrix from the historical time to the current time and the heat flux matrix at the next time step, respectively, to extract high-dimensional temperature field features that incorporate spatiotemporal context information. High-dimensional heat flux characteristics ; Multi-head cross-attention physical guidance module, including Each individual attention point Parallel computation of independent attention heads to high-dimensional temperature field features Through the first linear mapping matrix Projected into a query matrix High-dimensional heat flux characteristics Through the second linear mapping matrix and the third linear mapping matrix Projection bonding matrices respectively Sum matrix Each attention head outputs high-dimensional heat flux features. Characteristics of high-dimensional temperature fields The dynamic guidance response is obtained by splicing the outputs of all attention heads and performing a linear transformation to obtain the heat flux guidance feature. The feature fusion module combines heat flux guiding features with the original high-dimensional temperature field features. Perform adaptive fusion to obtain fused features ; Decoder, Input Fusion Features The predicted temperature field for the next time step is reconstructed. ; Step 3: Based on the temperature field predicted in Step 2, the next time step is determined. Extract the surface width of the molten pool With length The height of the molten pool is obtained by inputting a pre-trained geometric mapping model. With melting depth Then, the real-time volume, real-time dilution rate, and relative volume change rate relative to the statistically stable volume are calculated. Step 4: Use the relative volume change rate and dilution rate deviation as inputs to the fuzzy controller. The fuzzy controller outputs a laser power compensation command to drive the laser to adjust its power, thereby achieving dynamic closed-loop control of the repair process.

[0007] Furthermore, in step 1, the heat flux matrix and the temperature field matrix have the same size, and their matrix elements... The calculation formula is: In the formula, For the element indices of the heat flux matrix, The center index coordinates of the scan path points in the matrix. For laser power, For absorption rate, Where is the radius of the light spot. This represents the actual physical size corresponding to a single matrix unit.

[0008] Furthermore, in step 2, the dual-branch feature encoder module includes: a temperature field encoder. The ConvLSTM has a greater number of recursive layers than the thermal flux encoder. The number of recursive layers; where the thermal flux encoder Temperature field encoder employs fewer layers to capture the high-frequency transient input characteristics of lasers. A higher number of layers are used to model the deep semantic features of the thermal accumulation effect of the temperature field.

[0009] Furthermore, the ConvLSTM uses a forget gate. Input gate Output gate and memory unit Update internal hidden state The updated equation is: In the formula, The input tensor at the current moment, For convolution operators, For Hadama accumulation, This represents the current internal state of the memory cell. This is the input gate, used to determine how much new information can enter the internal state of the memory cell at the current moment. This is the forget gate, used to determine how much information from the previous memory cell's state can be forgotten. This is an output gate used to determine how much information from the current memory cell's internal state can be output. It is a non-linear activation function. For activation functions; , , , For input tensors The convolution weight matrix to the input gate, forget gate, output gate, and the internal state of the memory unit. , , , The hidden state from the previous moment The convolution weight matrix to the input gate, forget gate, output gate, and the internal states of the memory unit. , , The internal state of the cell at the previous moment The convolution weight matrices to the input gate, forget gate, and output gate; , , , These are the bias vectors for the input gate, forget gate, internal state of the memory unit, and output gate, respectively.

[0010] Furthermore, in step 2, the training of the molten pool temperature field prediction model adopts a two-stage strategy of "simulation pre-training - experimental fine-tuning": in the simulation pre-training stage, the model is pre-trained using a simulation dataset with the physical constraint loss function containing the heat conduction residual loss as the optimization objective; in the experimental fine-tuning stage, based on the pre-training, the model is fine-tuned using an experimental dataset, and the neural network weight parameters and learnable residual gating parameters are updated simultaneously.

[0011] Furthermore, the outputs of all attention heads in step 2 are concatenated and subjected to a linear transformation to obtain the heat flux guiding features, including: in These are the features after fusion.

[0012] Furthermore, the total loss function of the molten pool temperature field prediction model in step 2... , in To predict the mean square error between the temperature field and the measured temperature field, For the physical residual loss based on the transient heat conduction equation, and These are the corresponding loss weight coefficients; The physical residual loss is discretized using an explicit forward difference scheme: In the formula, For material density, is a function of specific heat capacity. It is a function of thermal conductivity. Here is the temperature field matrix. For the equivalent heat source term, For time step, Temperature field matrix The Laplace term, The spatial gradient of thermal conductivity, This is the spatial gradient term.

[0013] Furthermore, the operation of calculating the real-time volume, dilution rate, and relative volume change rate relative to the statistically stable volume in step 3 is as follows: Substitute the predicted length and width of the molten pool into the DNN geometric mapping model to calculate the predicted height h and melting depth d at the current moment in real time; calculate the real-time volume using the instantaneous volume formula. Calculate the real-time dilution rate. ; Calculate the real-time volume relative to the statistically stable volume relative volume change rate ; The DNN geometric mapping model described therein is a three-layer fully connected network structure; This is a statistical mean of the test results from multiple sampling points.

[0014] Furthermore, the statistically stable volume The method for obtaining the data is as follows: during the process test stage, multiple single-pass additive manufacturing experiments are carried out for the material to be repaired by changing the laser power; metallographic observation is performed on the experimental samples to obtain measured data, and multiple sets of measured data are denoised and averaged to calculate the statistical stable volume under the process parameters.

[0015] Furthermore, in step 4, the fuzzy controller adopts a two-input single-output structure, with the input being the relative volume change rate. and dilution deviation Set the fuzzy universe of discourse for the input and output quantities, and linearly normalize them to [-1,1], and divide them into five fuzzy subsets: negative large (NB), negative small (NS), zero (ZE), positive small (PS), and positive large (PB). In this method, adaptive compensation and control of laser power is achieved by constructing a multi-input coupled fuzzy rule base.

[0016] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention proposes a mechanism-data fusion-driven real-time quality control method for additive repair. By modeling the dynamic guidance of the temperature field by the heat source through dual-branch feature extraction and multi-head cross-attention mechanism, it solves the problem of accurately predicting the temperature field in complex transient thermal processes and achieves high-precision prediction of the molten pool morphology.

[0017] 2. This invention incorporates the physical residuals of the heat conduction equation as constraints into the model loss function, and introduces the mechanism of material property changes with temperature into model training. This effectively solves the problems of weak generalization ability and physical inconsistency of pure data-driven models under complex working conditions, and significantly improves the physical rationality and accuracy of predictions.

[0018] 3. This invention employs a simulation pre-training-actual fine-tuning strategy to optimize the model and introduces learnable residual gating parameters for adaptive fusion, enhancing the model's adaptability. Furthermore, it utilizes the model's predicted temperature field to invert the molten pool geometry and employs a fuzzy controller to output power compensation commands in real time, achieving dynamic closed-loop quality control of the repair process. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall framework of the present invention; Figure 2 This is a temperature distribution diagram of the additive repair thermal simulation based on JAX-AM according to the present invention; Figure 3 This is a structural diagram of the physical information neural network of the present invention; where Encoder represents the encoder, Decoder represents the decoder, and gate represents the gate. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions in the embodiments 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, of the embodiments of this invention. The JAX-AM used in the embodiments of this invention is an application-layer software built on the JAX machine learning framework to solve physical simulation problems in the field of additive manufacturing, used to simulate physical phenomena in the 3D printing process. 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.

[0021] Example In this embodiment, the mechanism-data fusion-driven real-time quality control method for additive repair achieves high-precision real-time prediction and control of the molten pool quality during the repair process through a combination of physical information neural networks and fuzzy control. Figure 1 This is a schematic diagram of the overall framework of the present invention, which is specifically carried out according to the following steps: Step 1: Constructing experimental and simulation datasets Temperature field data of the repair area was obtained through measured temperature acquisition and finite element simulation, and the corresponding heat flux matrix was generated by combining the laser path planning parameters to form a spatiotemporally aligned dataset. The specific process is as follows: Real-time thermal images of the repaired area were acquired using an infrared camera to obtain the measured temperature field matrix. Finite element thermal simulation was performed based on the JAX-AM platform to obtain the simulated temperature field matrix. Based on the laser path planning parameters, calculate the preset heat flux matrix for the next time step. This matrix has the same size as the temperature field matrix, and its matrix elements... The calculation formula is: In the formula, For the element indices of the heat flux matrix, The center index coordinates of the scan path points in the matrix. For laser power, For absorption rate, Where is the radius of the light spot. This represents the actual physical size corresponding to a single matrix unit.

[0022] The measured temperature field matrix With the simulated temperature field matrix Compare with the same preset heat flux matrix calculated based on laser path planning parameters. (Spatial resolution 0.25mm) Spatiotemporal feature alignment is performed using a bilinear interpolation algorithm. In this process, the temperature field at time t+1 is determined by the temperature field data at time t and earlier, as well as the heat flux input at time t+1. During model training, q(t+1) is input to predict T(t+1). In real-time control scenarios, the controller pre-sets the future heat flux and superimposes it with the temperature field matrix at time t to obtain the predicted temperature field matrix at time t+1, forming the measured dataset and the simulation dataset. Finally, all aligned data serve as the unified input for the subsequent dual-branch feature extraction architecture.

[0023] In this embodiment, 300M steel was used, and finite element thermal simulation was performed using JAX-AM to experimentally set the absorption rate. The value is 0.35, and the spot radius is... The thickness is 1.75 mm, the scanning speed is 10 mm / s, and the physical size of the center grid of the defect area (i.e., the actual physical size corresponding to a single matrix unit) is... The thickness is 0.25 mm. The simulation employs a spatially non-uniform mesh strategy, locally refining the simulation mesh in the core region of the molten pool where temperature gradients are severe, while gradually thinning it out in the far field. A transition ratio of 1:3 is set in the X, Y, and Z directions to achieve a seamless transition from the microscopic molten pool to the macroscopic workpiece, meaning that the size of any given mesh cell cannot exceed three times its adjacent mesh cell size. This balances high gradient capture in the molten pool core region with overall computational efficiency. To cover the extreme thermophysical laws from low-energy non-fusion to high-energy overheating, a simulation dataset is created within the laser power range of 800W to 2500W, sampled in 100W increments. The corresponding temperature field distribution is shown below. Figure 2 As shown, real-time thermal images of the repair area were simultaneously acquired using an infrared camera within a power range of 1500W to 2200W in 50W increments. A measured dataset was then created to correct for thermal accumulation deviations under actual repair conditions.

[0024] Step 2: Predict the temperature field at the next moment using the molten pool temperature field prediction model. Input the temperature field matrix from the historical moment to the current moment and the heat flux matrix for the next moment from the simulation dataset constructed in step 1 into the trained molten pool temperature field prediction model to obtain the temperature field for the next moment. .

[0025] like Figure 3 As shown, the molten pool temperature field prediction model includes: a dual-branch feature encoder module, a multi-head cross-attention physical guidance module, a feature fusion module, and a decoder.

[0026] The dual-branch feature encoder module consists of a temperature field encoder. With thermal flux encoder The system is composed of two parts: a temperature field matrix from the historical time to the current time and a heat flux matrix for the next time step, which are input sequentially according to time sequence. ConvLSTM temporal recurrent convolution is performed on both to capture the spatial distribution pattern and the temporal evolution trend of the data, and to extract high-dimensional temperature field features that integrate spatiotemporal context information. High-dimensional heat flux characteristics .

[0027] ConvLSTM is a variant of the traditional Long Short-Term Memory (LSTM) network. LSTM has three gates: the input gate, the output gate, and the forget gate. ConvLSTM primarily changes the weights of W to be convolutional operations, thus extracting features from two-dimensional spatial images. By learning time-varying temperature field and heat flux matrices, ConvLSTM extracts features on a temporal scale and fuses the spatiotemporal contextual information of the temperature field and heat flux matrices.

[0028] ConvLSTM can fully extract features from multivariate spatiotemporal sequences and continuously learn the long-term dependencies between them. Specifically, ConvLSTM replaces matrix multiplication with convolution operations at each gate of an LSTM unit. This captures fundamental spatial features by performing convolution operations on multidimensional data. The main difference between ConvLSTM and LSTM lies in the input dimensionality. Because LSTM input data is one-dimensional, it is not suitable for spatial sequence data, such as temperature field matrices or heat flux matrices. ConvLSTM uses 3D data as input, allowing it to consider the spatiotemporal correlations within temperature field data.

[0029] In this embodiment, a thermal flux encoder is set. ConvLSTM recursive layer number Less than temperature field encoder Number of recursion levels .in, A smaller number of layers are used to achieve rapid feature capture of high-frequency transient laser inputs. A higher number of layers are used to model the deep semantic features of the thermal accumulation effect of the temperature field.

[0030] Specifically, temperature field encoder The ConvLSTM has 4 layers and is a thermal flux encoder. The ConvLSTM has 2 layers.

[0031] At each time step, ConvLSTM passes through the forget gate. Input gate Output gate and memory unit Update internal hidden state The updated equation is as follows: In the formula, The input tensor at the current moment, For convolution operators, For Hadama accumulation, This represents the current internal state of the memory cell. This is the input gate, used to determine how much new information can enter the internal state of the memory cell at the current moment. This is the forget gate, used to determine how much information from the previous memory cell's state can be forgotten. This is an output gate used to determine how much information from the current memory cell's internal state can be output. It is a non-linear activation function. For activation functions; , , , For input tensors The convolution weight matrix to the input gate, forget gate, output gate, and the internal states of the memory unit. , , , The hidden state from the previous moment The convolution weight matrix to the input gate, forget gate, output gate, and the internal states of the memory unit. , , The internal state of the cell at the previous moment The convolution weight matrices to the input gate, forget gate, and output gate; , , , These are the bias vectors for the input gate, forget gate, internal state of the memory unit, and output gate, respectively.

[0032] The dual-branch ConvLSTM encoder described above extracts high-dimensional temperature field features that incorporate spatiotemporal context information. High-dimensional heat flux characteristics This provides feature input for subsequent multi-head cross-attention physical guidance modules.

[0033] Multi-head cross-attention physical guidance module, including Each individual attention point Parallel computation of independent attention heads to high-dimensional temperature field features Through the first linear mapping matrix Projected into a query matrix High-dimensional heat flux characteristics Through the second linear mapping matrix and the third linear mapping matrix Projection bonding matrices respectively Sum matrix Each attention head outputs high-dimensional heat flux features. Characteristics of high-dimensional temperature fields The dynamic guided response is obtained by splicing the outputs of all attention heads and performing a linear transformation to obtain the heat flux guided feature.

[0034] In this embodiment, The number of heads to focus on The value is 8: The high-dimensional temperature field features extracted by the dual-branch feature encoder module at the i-th attention head. Through the first linear mapping matrix Projected into a query matrix High-dimensional heat flux features extracted by the dual-branch feature encoder module Through the second linear mapping matrix and the third linear mapping matrix Projection bonding matrices respectively Sum matrix ; In the formula, , and Used to project input features into different semantic subspaces.

[0035] Each independent attention head computes high-dimensional heat flux features in parallel. Characteristics of high-dimensional temperature fields The dynamic guided response, the output of each attention head is represented as: In the formula, the query matrix Multiply by the key matrix The transpose of is used to obtain the similarity score, which represents the correlation strength between all temperature points and all heat source points. This is a scaling factor used to reduce the similarity scores to a suitable range in order to stabilize the gradient distribution; This means normalizing each row of the similarity matrix so that the sum of each row is 1, thus converting the similarity score into attention weights. Used to extract high-dimensional heat flux features based on attention weights The value matrix of the mapping Extract heat information and update high-dimensional temperature field features. ; Indicates the first A query matrix with attention heads; Indicates the first The key matrix of each attention head; Indicates the first The value matrix of each attention head.

[0036] The outputs of all attention heads are concatenated and then subjected to a linear transformation to obtain the final multi-head attention output: In the formula, For multi-head feature splicing operators; The output linear transformation matrix is ​​used to map the concatenated features back to the original dimensions. The multi-head cross-attention output is used to characterize the physical guidance information of heat flux on the spatiotemporal evolution of the temperature field, i.e., heat flux guidance features, and serves as the input features for subsequent temperature field prediction or reconstruction modules.

[0037] The feature fusion module introduces learnable residual gating parameters. , output multi-head attention Features of the original high-dimensional temperature field Perform adaptive fusion to obtain fused features : In the formula, This is a learnable scalar parameter used to dynamically adjust the fusion strength of attention-guided information, enabling the model to adaptively balance the importance of the original high-dimensional temperature field features and the heat flux-guided features according to the current operating conditions. In this embodiment, it is initialized to 0.2.

[0038] Decoder, Input Fusion Features The predicted temperature field for the next time step is reconstructed. The decoder includes a front-end spatiotemporal feature extraction unit and an end-end spatial reconstruction output unit. The front-end spatiotemporal feature extraction unit consists of... A stacked ConvLSTM layer (in this embodiment, the number of layers in the decoder front-end ConvLSTM is set to 4) is used to extract spatiotemporal evolution features from the fused features; the end spatial reconstruction output unit is a Conv3D output head, used to reconstruct the spatial dimension of the spatiotemporal evolution features, and finally outputs the next time-step temperature field with the same size as the input temperature field matrix. .

[0039] Total loss function of the molten pool temperature field prediction model For data loss With physical residual loss Weighted sum: in, To predict the mean square error between the temperature field and the measured temperature field, The residual loss due to heat conduction is constructed based on the transient heat conduction equation, and its continuous form is as follows: In the formula, and For the corresponding loss weight coefficient (in this embodiment, Set to 0.7. (Set to 0.3) For material density, is a function of specific heat capacity. It is a function of thermal conductivity. Here is the temperature field matrix. This is the equivalent heat source term.

[0040] To achieve differentiable discretization, an explicit forward difference scheme is adopted: the density constant of the material is pre-established. Specific heat capacity function With thermal conductivity function (This process involves importing the chemical composition of 300M steel into JmatPro material property simulation software to calculate its thermophysical parameters), and then applying the current temperature field matrix. The instantaneous physical property parameters corresponding to each grid point in space are obtained through numerical dynamic mapping; the current temperature field matrix is ​​extracted using the central difference operator. Laplace item Spatial gradient term and the spatial gradient of thermal conductivity This is used to characterize the spatially non-uniform evolution of the temperature field. The temperature field at the next moment is predicted from the temperature field. Instantaneous physical properties, characteristics of partial derivatives, and time step Substituting into the explicit forward difference scheme, we obtain the discretized heat conduction residual loss: The molten pool temperature field prediction model is pre-trained using a two-stage training strategy: simulation pre-training followed by experimental fine-tuning. The molten pool temperature field prediction model is pre-trained using the simulation dataset constructed in step 1. The simulation data covers a wide parameter range from low-energy incomplete fusion to high-energy overheating, enabling the molten pool temperature field prediction model to learn temperature field evolution characteristics that conform to the physical laws of heat conduction. This stage uses a physically constrained loss function. To optimize the objective, update the network weight parameters.

[0041] Based on pre-training, the molten pool temperature field prediction model was fine-tuned using a real-world dataset. The real-world data, derived from thermal images of the actual restoration environment captured by an infrared camera, reflects complex conditions such as actual heat accumulation and boundary heat dissipation. During fine-tuning, the weight parameters of the neural network and the learnable residual gating parameters were updated simultaneously. This enables the molten pool temperature field prediction model to adaptively adjust the fusion intensity, better matching the dynamic characteristics of the temperature field during the actual repair process. The structure diagram of the physical information neural network is shown below. Figure 3 As shown.

[0042] Through the above two-stage training, the molten pool temperature field prediction model combines consistency with physical laws and adaptability to actual working conditions, providing a high-precision and robust temperature field prediction foundation for subsequent online closed-loop control.

[0043] Step 3: Molten pool morphology inversion The temperature field at the next moment predicted in step 2 Extract the geometric length and width of the molten pool surface The internal morphological features of the molten pool are obtained by inputting the trained DNN geometric mapping model. Then, the real-time volume is calculated synchronously using the instantaneous volume formula. Real-time dilution rate and real-time volume Relative to statistical steady volume relative volume change rate .

[0044] The training steps for the DNN geometric mapping model include: conducting multiple sets of single-pass additive manufacturing experiments on the material to be repaired by varying the laser power; performing cross-sectional slicing and metallographic observation on each experimental sample to obtain the surface width of the molten pool. With length Corresponding molten pool height With melting depth The experimental data was used. Noise and mean values ​​were applied to multiple sets of experimental data to eliminate random disturbances. The experimental data was then divided into training and validation datasets. The training dataset was input into a DNN geometric mapping model built on a deep neural network for training. The model input was normalized using Min-Max, and the output was restored using Min-Max inverse normalization. The ReLU activation function was used for the hidden layers, and the output layer was linear with no activation. Mean squared error (MSE) was used as the loss function for model training, and the dataset was randomly divided into training and test sets in a 7:3 ratio.

[0045] The DNN geometric mapping model employs a three-layer fully connected network structure, based on the geometric dimensions of the molten pool surface. As input, the internal morphological features of the molten pool This is the output.

[0046] During the process testing phase, a stable forming process is selected, and the instantaneous volume of multiple sampling points within this phase is calculated: The statistically stable volume under this process parameter is obtained by statistically averaging the volumes of n sampling points during the steady-state phase. .

[0047] It should be noted that the statistically stable volume cannot be obtained directly by averaging the volumes of sampling points during real-time repair. This is because a thermal accumulation effect exists during actual repair, causing the statistically average volume of the repaired area to gradually increase as the process continues, resulting in a drift of the standard volume, which cannot be used as a fixed reference benchmark. Therefore, the statistically stable volume must be determined in advance during the process testing phase.

[0048] In the actual repair process, the predicted length and width of the molten pool are substituted into the established DNN geometric mapping model to calculate the predicted height at the current moment in real time. With melting depth The instantaneous volume formula is used to simultaneously calculate the real-time volume. And calculate the real-time dilution rate: Further calculation of real-time volume Relative to statistical steady volume relative volume change rate : Step 4: Fuzzy Control The fuzzy controller is based on the relative volume change rate. and the incremental output laser power compensation for dilution rate deviation .

[0049] Specifically, it includes: Step 41. Set the relative volume change rate The range is [-5%, 5%], the dilution rate deviation range is [-8%, 8%], and the laser power compensation increment is [-5%, 5%]. The range is [-200W, +200W], and it is linearly normalized to [-1, 1].

[0050] Step 42. Calculate the relative volume change rate. and dilution deviation Laser power compensation increment Each is divided into five fuzzy subsets {negative large, negative, zero, positive, positive large}, corresponding to records {NB, NS, ZE, PS, PB}; Step 43. Input fuzzification: Fuzzify the input quantity using the relative volume change rate. And dilution rate deviation fuzzification to obtain membership degrees on different fuzzy subsets: Calculate the relative volume change rate Membership degree on 5 fuzzy subsets Calculate dilution deviation Membership degree on 5 fuzzy subsets Step 44. Perform fuzzy inference on the membership degrees of the input quantities according to the set fuzzy rules to obtain the laser power compensation increment. The corresponding fuzzy subset; Table 1 Fuzzy Rule Table In this embodiment, the input variable and Output variables The fuzzy subsets are defined by triangular membership functions, covering their respective domains. Using the Mamdani fuzzy inference method, the membership degrees of the two input variables on their respective five fuzzy subsets are first minified according to the fuzzy rule table in Table 1, yielding the trigger strengths corresponding to 25 rules; then, for the same... For multiple rules in a fuzzy subset, their trigger strengths are maximized (max) to aggregate the rules; subsequently, the aggregated trigger strengths are used to apply the corresponding rules... The membership functions of the fuzzy subsets are truncated, and multiple output fuzzy sets are aggregated into a maximal standard union to form the final output fuzzy set. Finally, the discrete centroid method is used to defuzzify the output fuzzy set to obtain a clear set. Output value.

[0051] In terms of specific control logic, when detected When the relative volume change rate is positive, the fuzzy inference engine outputs a negative power compensation increment. This is achieved by reducing the laser energy power to suppress excessively rapid expansion of the molten pool; simultaneously, if the dilution rate deviates... A negative value (indicating that the actual dilution rate is lower than the target value and the substrate melt depth is insufficient) will cause the fuzzy controller to reduce the negative compensation amplitude to ensure the melt depth and the bonding strength between the repair layer and the substrate, thus ensuring the coordinated stability of the forming morphology and the dilution rate.

[0052] The resulting power compensation increment The laser is driven by superimposed power from the base process to achieve dynamic closed-loop correction of the molten pool volume and forming quality during the repair process. The preset heat flux matrix q(t+1) changes with the power variable. The new temperature field and the heat flux matrix changing in the next time step predict the relative volume change rate in the next step. And the input amount of dilution rate deviation.

[0053] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A mechanism-data fusion-driven real-time quality control method for additive repair, characterized in that, Includes the following steps: Step 1: Obtain the measured temperature field matrix of the repair area through infrared thermal imaging, and generate the corresponding heat flux matrix by combining the laser path planning parameters. Align the temperature field matrix and the heat flux matrix in terms of spatiotemporal features to construct a training dataset. Step 2: Input the temperature field matrix from the historical time to the current time and the heat flux matrix for the next time into the pre-trained molten pool temperature field prediction model to obtain the temperature field for the next time. The molten pool temperature field prediction model includes: a dual-branch feature encoder module, a multi-head cross-attention physical guidance module, a feature fusion module, and a decoder; The dual-branch feature encoder module consists of a temperature field encoder. With thermal flux encoder The process involves performing ConvLSTM temporal recurrent convolution on the temperature field matrix from the historical time to the current time and the heat flux matrix at the next time step, respectively, to extract high-dimensional temperature field features that incorporate spatiotemporal context information. High-dimensional heat flux characteristics ; Multi-head cross-attention physical guidance module, including Each individual attention point Parallel computation of independent attention heads to high-dimensional temperature field features Through the first linear mapping matrix Projected into a query matrix High-dimensional heat flux characteristics Through the second linear mapping matrix and the third linear mapping matrix Projection bonding matrices respectively Sum matrix Each attention head outputs high-dimensional heat flux features. Characteristics of high-dimensional temperature fields The dynamic guidance response is obtained by splicing the outputs of all attention heads and performing a linear transformation to obtain the heat flux guidance feature. The feature fusion module combines heat flux guiding features with the original high-dimensional temperature field features. Perform adaptive fusion to obtain fused features ; Decoder, Input Fusion Features The predicted temperature field for the next time step is reconstructed. ; Step 3: Based on the temperature field predicted in Step 2, the next time step is determined. Extract the surface width of the molten pool With length The height of the molten pool is obtained by inputting a pre-trained geometric mapping model. With melting depth Then, the real-time volume, real-time dilution rate, and relative volume change rate relative to the statistically stable volume are calculated. Step 4: Use the relative volume change rate and dilution rate deviation as inputs to the fuzzy controller. The fuzzy controller outputs a laser power compensation command to drive the laser to adjust its power, thereby achieving dynamic closed-loop control of the repair process.

2. The real-time quality control method for additive repair based on mechanism-data fusion as described in claim 1, characterized in that, In step 1, the heat flux matrix and the temperature field matrix have the same size and their matrix elements are... The calculation formula is: In the formula, For the element indices of the heat flux matrix, The center index coordinates of the scan path points in the matrix. For laser power, For absorption rate, Where is the radius of the light spot. This represents the actual physical size corresponding to a single matrix unit.

3. The real-time quality control method for additive repair based on mechanism-data fusion as described in claim 1, characterized in that, In step 2, the dual-branch feature encoder module includes: a temperature field encoder. The ConvLSTM has a greater number of recursive layers than the thermal flux encoder. The number of recursion levels; Among them, thermal flux encoder Temperature field encoder employs fewer layers to capture the high-frequency transient input characteristics of lasers. A higher number of layers are used to model the deep semantic features of the thermal accumulation effect of the temperature field.

4. The real-time quality control method for additive repair based on mechanism-data fusion as described in claim 3, characterized in that, The ConvLSTM uses a forget gate. Input gate Output gate and memory unit Update internal hidden state The updated equation is: In the formula, The input tensor at the current moment, For convolution operators, For Hadama accumulation, This represents the current internal state of the memory cell. This is the input gate, used to determine how much new information can enter the internal state of the memory cell at the current moment. The forget gate determines how much information from the previous memory cell's state can be forgotten. This is an output gate used to determine how much information from the current memory cell's internal state can be output. It is a non-linear activation function. For activation functions; , , , For input tensors The convolution weight matrix to the input gate, forget gate, output gate, and the internal states of the memory unit. , , , The hidden state from the previous moment The convolution weight matrix to the input gate, forget gate, output gate, and the internal states of the memory unit. , , The internal state of the cell at the previous moment The convolution weight matrices for the input gate, forget gate, and output gate; , , , These are the bias vectors for the input gate, forget gate, internal state of the memory unit, and output gate, respectively.

5. The real-time quality control method for additive repair based on mechanism-data fusion as described in claim 1, characterized in that, In step 2, the training of the molten pool temperature field prediction model adopts a two-stage strategy of simulation pre-training and experimental fine-tuning: in the simulation pre-training stage, the model is pre-trained using a simulation dataset, with the physical constraint loss function, which includes the heat conduction residual loss, as the optimization target. In the experimental fine-tuning phase, based on the pre-training, the model is fine-tuned using the experimental dataset, and the neural network weight parameters and learnable residual gating parameters are updated simultaneously.

6. The real-time quality control method for additive repair based on mechanism-data fusion as described in claim 1, characterized in that, In step 2, the outputs of all attention heads are concatenated and subjected to linear transformation to obtain the heat flux guiding features, including: in These are the features after fusion.

7. The real-time quality control method for additive repair based on mechanism-data fusion as described in claim 1, characterized in that, The total loss function of the molten pool temperature field prediction model in step 2 , in To predict the mean square error between the temperature field and the measured temperature field, The physical residual loss is based on the transient heat conduction equation. and These are the corresponding loss weight coefficients; The physical residual loss is discretized using an explicit forward difference scheme: In the formula, For material density, is a function of specific heat capacity. It is a function of thermal conductivity. Here is the temperature field matrix. For the equivalent heat source term, For time step, Temperature field matrix The Laplace term, The spatial gradient of thermal conductivity, This is the spatial gradient term.

8. The real-time quality control method for additive repair based on mechanism-data fusion as described in claim 1, characterized in that, The operation of calculating the real-time volume, dilution rate, and relative volume change rate relative to the statistically stable volume in step 3 is as follows: substitute the predicted length and width of the molten pool into the DNN geometric mapping model, and solve the predicted height h and melting depth d corresponding to the current moment in real time. Calculate real-time volume using the instantaneous volume formula Calculate the real-time dilution rate. ; Calculate the real-time volume relative to the statistically stable volume relative volume change rate ; The DNN geometric mapping model described therein is a three-layer fully connected network structure; This is a statistical mean of the test results from multiple sampling points.

9. The real-time quality control method for additive repair based on mechanism-data fusion as described in claim 8, characterized in that, The statistically stable volume The method for obtaining the data is as follows: during the process test stage, multiple single-pass additive manufacturing experiments are carried out for the material to be repaired by changing the laser power; metallographic observation is performed on the experimental samples to obtain measured data, and multiple sets of measured data are denoised and averaged to calculate the statistical stable volume under the process parameters.

10. The real-time quality control method for additive repair based on mechanism-data fusion as described in claim 1, characterized in that, In step 4, the fuzzy controller adopts a two-input single-output structure, with the input being the relative volume change rate. and dilution deviation Set the fuzzy universe of discourse for the input and output quantities, and linearly normalize them to [-1,1], and divide them into five fuzzy subsets: negative large NB, negative small NS, zero ZE, positive small PS, and positive large PB. In this method, adaptive compensation and control of laser power is achieved by constructing a multi-input coupled fuzzy rule base.

Images