Artificial intelligence-based path planning method and system for repairing surface damage of metal parts by ultra-micro 3D printing

Abstract

The application relates to a path planning method and system for repairing surface damage of a metal part by super-micro 3D printing based on artificial intelligence, which comprises the following steps: detecting damage of a metal surface image by using an improved YOLO network embedded with a coordinate attention mechanism and a minimum damage detection layer, combining point cloud data reconstruction to generate a three-dimensional damage model; performing self-adaptive layered slicing based on damage depth distribution and generating a repair slice data set; mapping the slice data into a coloring traveling salesman problem, decomposing a sub-problem by K-means clustering and solving and optimizing a deposition path by using a simulated annealing-variable neighborhood search hybrid algorithm; constructing a multi-physical field coupling model, taking the optimized path as a constraint to perform numerical simulation, using simulation data to train an improved long short-term memory network to predict process parameters of each path segment; collecting state data in real time through a sensor, fusing a predicted value and a measured value by using extended Kalman filtering and dynamically correcting a control parameter; and identifying a repair abnormal type and calling a corresponding repair strategy.

Description

Technical Field

[0001] This application relates to the field of 3D printing technology, and in particular to a path planning method and system for repairing surface damage of metal parts using artificial intelligence-based micro 3D printing. Background Technology

[0002] Metal components serve in high-end equipment such as aero engines, gas turbines, precision molds, and marine transmission systems for extended periods. Their surfaces inevitably suffer damage such as cracks, pitting, scratches, and peeling due to wear, corrosion, fatigue, and impact. These critical components are extremely expensive to manufacture, with individual pieces costing hundreds of thousands or even millions of yuan. Replacing the entire component due to surface damage would result in enormous economic waste.

[0003] Ultra-micro 3D printing technology deposits metal powder layer by layer onto the damaged area with micron-level precision, enabling precise repair by filling only the missing parts. It balances repair accuracy with metallurgical bonding quality and is considered an ideal technology for repairing surface damage in metal parts. However, the DED (Deposit-Effect 3D) repair process involves multiple stages, including damage morphology perception, deposition path planning, process parameter control, and online quality monitoring. These stages are complexly coupled, and deviations in any stage can lead to repair failure or substandard repair quality.

[0004] Currently, the repair of metal parts still relies on operators visually judging the extent and depth of damage using simple tools. This is not only time-consuming and labor-intensive, but also makes it easy to miss extremely small damages such as microcracks and micro-corrosion. Damage depth often depends on estimation, lacking precise three-dimensional geometric reconstruction, which cannot provide an accurate digital model for subsequent repairs. This can easily lead to rough determination of the repair area boundary and excessive or insufficient repair margin.

[0005] Based on this, this application provides a path planning method and system for repairing surface damage of metal parts using artificial intelligence-based micro 3D printing. Summary of the Invention

[0006] To address the issues of easily overlooking extremely small damages such as microcracks and micro-corrosion, and the fact that damage depth often relies on estimation and lacks precise three-dimensional geometric reconstruction, thus failing to provide accurate digital models for subsequent repairs and easily leading to coarse determination of repair area boundaries and excessive or insufficient repair margins, this application provides a path planning method and system for repairing surface damage of metal parts using artificial intelligence-based ultra-micro 3D printing.

[0007] Firstly, this application provides a path planning method for repairing surface damage of metal parts using ultra-micro 3D printing based on artificial intelligence, employing the following technical solution: including: An improved YOLO detection network is used to perform damage detection on the surface images of metal parts. The detection results are mapped to depth point cloud data, and a three-dimensional damage model is generated through region growing and Poisson surface reconstruction. Based on the three-dimensional damage model, the repair area is determined, adaptive layered slicing is performed, and a repair slice dataset is generated. The repaired slice dataset is mapped to the Colored Traveling Salesman Problem. A partially optimized metaheuristic decomposition framework is used to decompose the subproblem through K-means clustering, and a simulated annealing-variable neighborhood search hybrid heuristic algorithm is used to solve it. After path repair and greedy integration, the optimal deposition path sequence is obtained. A multiphysics coupling model for metal directional energy deposition was constructed, and numerical simulation was performed with the optimal deposition path sequence as spatial constraints. An improved long short-term memory network with added batch normalization and discard layers was trained using the simulation data to predict the optimal combination of process parameters for each path segment and obtain the optimized process parameter package. Real-time acquisition of repair status data; use extended Kalman filtering to fuse the predicted values ​​of the improved long short-term memory network with the measured values; when the deviation exceeds a preset threshold, trigger status correction and obtain correction control parameters. When a repair anomaly is detected, the corresponding repair strategy is invoked based on the correction control parameters to perform the anomaly repair.

[0008] Preferably, the step of using an improved YOLO detection network to perform damage detection on the surface image of the metal component, mapping the detection results to depth point cloud data, and generating a three-dimensional damage model through region growing and Poisson surface reconstruction includes: The improved YOLO detection network consists of a backbone network, a neck network, and a detection head. A coordinate attention mechanism module is inserted before the fast spatial pyramid pooling module in the backbone network. This module decomposes channel attention into one-dimensional feature encoding in the horizontal direction and one-dimensional feature encoding in the vertical direction, and aggregates features along the two spatial directions to generate an attention weight matrix. A high-pixel minimal damage detection layer is added to the neck network, expanding the detection layer to four levels. The improved YOLO detection network uses an EIOU loss function based on four geometric factors: overlapping region, center point distance, width difference, and height difference, for bounding box regression. Images of metal parts are captured using industrial cameras under multiple angles and lighting conditions. The surface image of the metal part is input into the improved YOLO detection network to obtain a two-dimensional bounding box of damage and damage classification results. The damage classification results include at least one of cracks, wear pits, pitting, scratches, spalling and oxidation corrosion pits. The two-dimensional bounding box of the damage is mapped to the depth point cloud data obtained by structured light or laser line scanning to determine the three-dimensional boundary of the damaged area; Based on the three-dimensional boundary of the damaged region, a region growing algorithm is used to extract the complete point cloud of the damaged region from the depth point cloud data. The Poisson surface reconstruction algorithm is used to reconstruct the surface of the complete point cloud to generate the damaged 3D model.

[0009] Preferably, the step of determining the repair area based on the three-dimensional damage model, performing adaptive layered slicing, and generating a repair slice dataset includes: The maximum depth of the damage is calculated based on the three-dimensional damage model. The repairability of the damage to be repaired is judged based on the preset repair feasibility criteria. The preset repair feasibility criteria include that the maximum depth of the damage does not exceed a threshold proportion of the component wall thickness, which is repairable; otherwise, it is not repairable. For damage that is determined to be repairable, a transition zone is expanded outward based on the three-dimensional boundary of the damage three-dimensional model to obtain a repair area; the transition zone includes a constant width zone near the damage side and a gradient width zone away from the damage side, and the width of the gradient width zone decreases to zero from the inside to the outside. Based on the damage depth distribution of the three-dimensional damage model, the repair area is divided into a bottom bonding sub-region, a main body filling sub-region, and a surface transition sub-region. The bottom bonding sub-region is sliced ​​with a first layer thickness, the main filling sub-region is sliced ​​with a second layer thickness, and the surface transition sub-region is sliced ​​with a decreasing layer thickness; the first layer thickness is less than the second layer thickness. A parallel adaptive layered slicing algorithm is used to synchronously perform slicing processing on each of the repair areas to generate slice contour data for each layer; the exposed area area and the maximum fill distance are calculated on the slice contour data, and the repair slice dataset is generated based on the slice contour data, the exposed area area and the maximum fill distance.

[0010] Preferably, the step of mapping the repaired slice dataset to a colored traveling salesman problem, using a partially optimized metaheuristic decomposition framework to decompose subproblems through K-means clustering, and solving them using a simulated annealing-variable neighborhood search hybrid heuristic algorithm, and obtaining the optimal deposition path sequence after path repair and greedy integration, includes: The deposition filling units in the repair slice dataset are mapped to cities in the Colored Traveling Salesman Problem, the repair nozzles are mapped to merchants, and the process constraint types are mapped to colors; the process constraint types include overlap ratio constraints, thermal input uniformity constraints, scan direction constraints, and partition transition constraints. With the goal of minimizing the weighted sum of total empty travel distance and heat accumulation penalty, a composite cost objective function for the colored traveling salesman problem is constructed. In this function, cities of the same color can only be visited by merchants carrying the same color, each city is visited exactly once, and each merchant's path forms an open-loop Hamiltonian path. A partially optimized metaheuristic decomposition framework is adopted, and K-means clustering is used to spatially cluster all the deposition filling units of the current slice layer to obtain several clusters; the dominant process constraint type of each cluster is determined as the main color; For each cluster, the deposition filling unit corresponding to its primary color is used as the core, and the deposition filling units located at the boundary of the adjacent clusters are introduced to construct an open-loop coloring traveling salesman problem model with a global perspective. A hybrid heuristic algorithm of simulated annealing and variable neighborhood search is used to solve each of the open-loop coloring traveling salesman problem models. The simulated annealing-variable neighborhood search hybrid heuristic algorithm employs a dual-chromosome encoding method, where the first chromosome marks the access order of the deposition filling units, and the second chromosome marks the number of the repair nozzle performing the deposition. In the initialization phase, a nearest-neighbor greedy strategy is used to generate an initial solution. In the neighborhood search phase, the inversion operator, the exchange operator, and the gene fragment re-insertion operator are used alternately. The solution is updated using the simulated annealing Metropolis criterion. Greedy integration of solutions to all the open-loop colored traveling salesman subproblem models yields an initial complete path sequence; The initial complete path sequence is repaired by checking time window constraints, modular constraints, corner constraints and thermal accumulation constraints. Cooling waiting time is inserted for path segments that violate the thermal accumulation constraints to obtain the optimal deposition path sequence.

[0011] Preferably, the construction of the metal-directed energy deposition multiphysics coupling model involves numerical simulation using the optimal deposition path sequence as a spatial constraint. An improved long short-term memory network with added batch normalization and discard layers is trained using the simulation data to predict the optimal combination of process parameters for each path segment, thereby obtaining an optimized process parameter package, including: A multiphysics coupling model for metal directional energy deposition is constructed; the multiphysics coupling model for metal directional energy deposition includes a laser energy distribution model, a molten pool temperature field model, and a molten pool fluid dynamics model. The laser energy distribution model uses a Gaussian surface heat source model to describe the energy distribution of the laser beam on the deposition surface; the molten pool temperature field model establishes a three-dimensional transient temperature field control equation based on Fourier's heat conduction law; the molten pool fluid dynamics model satisfies the Navier-Stokes equations of mass conservation and momentum conservation, and the driving force includes buoyancy driving terms and Marangoni convection terms. Using the optimal deposition path sequence as a spatial constraint, the sequence and position information of each path segment in the optimal deposition path sequence are transformed into the spatiotemporal trajectory of the moving heat source; the spatiotemporal trajectory is loaded into the metal directional energy deposition multiphysics coupling model through a user-defined function for numerical simulation; the numerical simulation uses fine mesh for the deposition channel and thermally affected zone, and coarse mesh for areas far from the deposition zone; The process status data for each path segment is extracted from the output of the numerical simulation; the process status data includes the maximum temperature of the molten pool, the cooling rate, the peak residual stress, and the thickness of the deposited layer; The process state data is used to construct a training dataset to train an improved long short-term memory network. A batch normalization layer is placed between the output of the gating unit of the original long short-term memory network and the input of the next layer to prevent gradient vanishing. A dropout layer randomly drops neurons during training at a preset dropout rate to prevent overfitting. The slice data of the model to be repaired and the optimal deposition path sequence are input into the trained improved long short-term memory network to predict the optimal process parameter combination for each path segment layer by layer; the optimal process parameter combination includes laser power, scanning speed, powder feeding rate and protective gas flow rate. If the maximum temperature of the molten pool predicted by the improved long short-term memory network exceeds a preset proportional threshold of the material's flash point, a cooling waiting strategy is automatically inserted in the corresponding path segment; the cooling waiting strategy includes at least one of extending the black screen time and increasing the protective gas flow rate. Based on the optimal combination of process parameters for each path segment and the cooling waiting strategy, the optimized process parameter package is generated.

[0012] Preferably, the real-time acquisition and repair status data uses an extended Kalman filter to fuse the predicted values ​​and measured values ​​of the improved long short-term memory network. When the deviation exceeds a preset threshold, a status correction is triggered, and correction control parameters are obtained, including: The repair status data is collected in real time using a molten pool vision camera, an infrared thermal imager, a laser displacement sensor, and an acoustic emission sensor. The molten pool vision camera collects molten pool morphology images at a first acquisition frequency, the infrared thermal imager collects molten pool temperature field images at a second acquisition frequency, the laser displacement sensor collects deposition layer thickness data at a third acquisition frequency, and the acoustic emission sensor collects acoustic emission signals at a fourth acquisition frequency. The process status data is used as the predicted value of the extended Kalman filter, and the repair status data is used as the measured value of the extended Kalman filter. Define the state variables and control inputs for the extended Kalman filter; the state variables include the melt pool temperature, the deposited layer thickness, and the cooling rate; the control inputs include the laser power correction and the scanning speed correction. Using the improved long short-term memory network as the state transition function, the predicted value of the state variable at the current time and the prediction error covariance are calculated based on the estimated value of the state variable at the previous time and the control input at the current time. Using the molten pool temperature measured by an infrared thermal imager and the deposition layer thickness measured by a laser displacement sensor as observation variables, the Kalman gain is calculated. The state variable is then corrected using the Kalman gain, the observation variables, and the predicted value to obtain the corrected state variable, and the prediction error covariance is updated. Calculate the temperature deviation between the corrected state variables and the maximum temperature of the melt pool predicted by the improved long short-term memory network, and calculate the height deviation between the corrected state variables and the target deposit thickness. When the temperature deviation exceeds a preset temperature threshold or the height deviation exceeds a preset proportional threshold for the target deposition layer thickness, a state correction is triggered. Based on the corrected state variable, the laser power and scanning speed of the corresponding path segment in the optimized process parameter package are adjusted to obtain the corrected control parameters.

[0013] Preferably, when a repair anomaly is detected, the step of invoking the corresponding repair strategy to perform anomaly repair based on the correction control parameters further includes: Based on the molten pool morphology image, deposition layer thickness data, molten pool temperature field image, and acoustic emission signal in the repair status data, the repair anomaly type is identified; the repair anomaly types include: lack of fusion, excessive deposition, surface cracks, spatter, porosity, and excessive heat accumulation. Based on the type of repair anomaly, the corresponding repair strategy is selected from a preset set of repair strategies; the set of repair strategies includes: local remelting strategy, in-situ subtractive material correction strategy, crack filling strategy, surface cleaning and re-deposition strategy, and forced cooling strategy; When the repair anomaly type is non-fusion, the local remelting strategy is invoked; the local remelting strategy includes: increasing the laser power based on the correction control parameters, decreasing the scanning speed based on the correction control parameters, and performing secondary scanning deposition on the path segment corresponding to non-fusion. When the repair anomaly type is over-deposition, the in-situ subtractive correction strategy is invoked; the in-situ subtractive correction strategy includes: using the height deviation of the deposition layer thickness in the correction control parameters as the cutting amount, invoking the accompanying subtractive tool to perform in-situ height correction on the path segment corresponding to over-deposition, and then performing deposition using the correction control parameters; When the repair anomaly type is a surface crack, a crack filling strategy is invoked; the crack filling strategy includes: reducing the scanning speed based on the correction control parameters, adjusting the powder feeding rate with the correction control parameters, and performing crack filling deposition along the crack direction; When the repair anomaly type is splash or pore, the surface cleaning and re-deposition strategy is invoked; the surface cleaning and re-deposition strategy includes: invoking an online polishing tool to clean the path segment corresponding to the splash or pore, and performing fixed-point re-deposition at the cleaned position using the correction control parameters; When the repair anomaly type is thermal accumulation exceeding the limit, a forced cooling strategy is invoked; the forced cooling strategy includes: inserting a cooling waiting time into the path segment corresponding to the thermal accumulation exceeding the limit, and adjusting the protective gas flow rate based on the correction control parameters.

[0014] Secondly, this application discloses a path planning device for repairing surface damage of metal parts using ultra-micro 3D printing based on artificial intelligence, which adopts the following technical solution, including: The 3D damage module is used to perform damage detection on surface images of metal parts using an improved YOLO detection network, map the detection results to depth point cloud data, and generate a 3D damage model through region growing and Poisson surface reconstruction. The repair data module is used to determine the repair area based on the damage 3D model, perform adaptive layered slicing, and generate a repair slice dataset. The deposition path module is used to map the repair slice dataset into a coloring traveling salesman problem. It adopts a partially optimized metaheuristic decomposition framework to decompose the subproblem through K-means clustering and solves it using a simulated annealing-variable neighborhood search hybrid heuristic algorithm. After path repair and greedy integration, the optimal deposition path sequence is obtained. The optimization process module is used to construct a multiphysics coupling model for metal directional energy deposition, perform numerical simulation with the optimal deposition path sequence as a spatial constraint, train an improved long short-term memory network with added batch normalization and discard layers using simulation data, predict the optimal combination of process parameters for each path segment, and obtain an optimized process parameter package. The parameter correction module is used to collect repair status data in real time. It uses extended Kalman filtering to fuse the predicted value and measured value of the improved long short-term memory network. When the deviation exceeds the preset threshold, the status correction is triggered to obtain the correction control parameters. The repair control module is used to call the corresponding repair strategy to perform abnormal repair based on the correction control parameters when a repair anomaly is detected.

[0015] Thirdly, this application also provides a control device, the device comprising: It includes a memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and executed, such as the above-described path planning method for repairing surface damage of metal parts using artificial intelligence-based micro 3D printing.

[0016] Fourthly, this application also provides a computer-readable storage medium storing a computer program that can be loaded by a processor and executed as described above regarding the path planning method for repairing surface damage of metal parts using artificial intelligence-based micro 3D printing.

[0017] In summary, the system in this application employs an embedded coordinate attention mechanism, adds a minimal damage detection layer, and uses an improved YOLO detection network with an EIOU loss function to detect damage on the surface images of metal parts. The detection results are mapped to depth point cloud data, and a 3D damage model is generated through region growing and Poisson surface reconstruction. Based on the 3D damage model, the repair area is determined, and adaptive layering is performed on the bottom bonding area, the main filling area, and the surface transition area according to the damage depth distribution, using different layer thicknesses to generate a repair slice dataset. The repair slice dataset is mapped to a colored traveling salesman problem, and sub-problems are divided using K-means clustering and solved using a hybrid heuristic algorithm of simulated annealing-variable neighborhood search. The optimal deposition path sequence is obtained through path repair and greedy integration. A multiphysics coupled model for metal-oriented energy deposition (MED) was constructed, incorporating laser energy distribution, molten pool temperature field, and molten pool fluid dynamics. Numerical simulations were performed using the optimal deposition path sequence as spatial constraints. An improved long short-term memory (LSTM) network with batch normalization and discard layers was trained using simulation data to predict the optimal combination of process parameters for each path segment, forming an optimized process parameter package. During the repair process, repair status data was collected in real-time using multi-source sensors. An extended Kalman filter was used to fuse the predicted values ​​from the improved LSD network with the measured values. When the deviation exceeded a preset threshold, a status correction was triggered, and corrected control parameters were obtained. When a repair anomaly was detected, the corresponding repair strategy was invoked based on the corrected control parameters for online anomaly repair. This approach achieved high-precision localization and 3D reconstruction of surface damage in metal parts, efficient planning of deposition paths, real-time dynamic optimization of process parameters, and online repair of various anomalies. It significantly improved repair accuracy and efficiency, greatly reduced the proportion of empty paths, and minimized process parameter prediction errors, providing a complete technical solution for ultra-micro 3D printing repair of surface damage in metal parts. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating a path planning method for repairing surface damage of metal parts using artificial intelligence-based micro 3D printing.

[0019] Figure 2 This is a structural block diagram of a path planning device for repairing surface damage of metal parts using artificial intelligence-based micro 3D printing. Detailed Implementation

[0020] The following combination Figures 1-2 This application will be described in further detail.

[0021] Reference Figure 1The embodiments of this application include at least steps S10 to S60.

[0022] S10 uses an improved YOLO detection network to perform damage detection on the surface image of the metal part, maps the detection results to depth point cloud data, and generates a three-dimensional damage model through region growing and Poisson surface reconstruction. S20: Based on the three-dimensional damage model, the repair area is determined, adaptive layered slicing is performed, and a repair slice dataset is generated. S30 maps the repair slice dataset to the colorized traveling salesman problem. A partially optimized metaheuristic decomposition framework is used to decompose the subproblem through K-means clustering, and a simulated annealing-variable neighborhood search hybrid heuristic algorithm is used to solve it. After path repair and greedy integration, the optimal deposition path sequence is obtained. S40, construct a multiphysics coupling model for metal directional energy deposition, perform numerical simulation with the optimal deposition path sequence as spatial constraint, use simulation data to train an improved long short-term memory network with added batch normalization layer and discard layer, predict the optimal combination of process parameters for each path segment, and obtain the optimized process parameter package; S50 collects repair status data in real time, uses extended Kalman filtering to fuse the predicted values ​​of the improved long short-term memory network with the measured values, and triggers status correction when the deviation exceeds the preset threshold to obtain correction control parameters; S60: When a repair anomaly is detected, the corresponding repair strategy is invoked based on the correction control parameters to perform the anomaly repair.

[0023] Specifically, the system improves the YOLO network to accurately reconstruct the 3D damage model, solving the problems of missing micro-defects and depth quantization; it uses depth partitioning and adaptive slicing to match the layer thickness with the damage gradient; it models the repair path planning as a colored traveling salesman problem, incorporates it into the process constraints through clustering decomposition and heuristic solution, significantly reducing empty running paths; it constructs a multiphysics model and trains an improved LSTM to quickly predict process parameters along the optimized path, replacing empirical settings and time-consuming simulations; it uses extended Kalman filtering to fuse multi-source sensing and model prediction, dynamically correcting parameters to eliminate accumulated bias; and it performs online local repairs on various anomalies based on the corrected parameters, avoiding overall rework.

[0024] In some embodiments, step S10 specifically includes the following steps: acquiring surface images of metal parts using an industrial camera under multi-angle and multi-light conditions; inputting the surface images of the metal parts into an improved YOLO detection network to obtain a two-dimensional bounding box of damage and damage classification results, wherein the damage classification results include at least one of cracks, wear pits, pitting, scratches, spalling, and oxidation corrosion pits; mapping the two-dimensional bounding box of damage to depth point cloud data obtained by structured light or laser line scanning to determine the three-dimensional boundary of the damaged area; based on the three-dimensional boundary of the damaged area, using a region growing algorithm to extract the complete point cloud of the damaged area from the depth point cloud data; and using a Poisson surface reconstruction algorithm to reconstruct the surface of the complete point cloud to generate a three-dimensional damage model.

[0025] The improved YOLO detection network includes a backbone network, a neck network, and a detection head. A coordinate attention mechanism module is inserted before the fast spatial pyramid pooling module of the backbone network. The coordinate attention mechanism module decomposes channel attention into one-dimensional feature encoding in the horizontal direction and one-dimensional feature encoding in the vertical direction, and aggregates features along the two spatial directions to generate an attention weight matrix. A high-pixel minimal damage detection layer is added to the neck network, expanding the detection layer to four levels. The improved YOLO detection network uses the EIOU loss function based on four geometric factors: overlapping region, center point distance, width difference, and height difference, for bounding box regression.

[0026] Specifically, the system acquires images of metal component surfaces from multiple angles and under various lighting conditions using an industrial camera, and inputs these images into an improved YOLO detection network. This network incorporates a coordinate attention mechanism module into its backbone, decomposing channel attention into one-dimensional feature encodings in both horizontal and vertical directions and generating attention weights. A high-resolution minimal damage detection layer is added, expanding the detection layer to four levels. An EIOU loss function based on four geometric factors—overlapping region, center point distance, width difference, and height difference—is used for bounding box regression, effectively improving the detection accuracy for extremely small-sized damage such as microcracks and micro-corrosion. After outputting the two-dimensional bounding box and classification results, the bounding box is mapped to depth point cloud data acquired by structured light or laser line scanning to determine the three-dimensional boundary of the damaged area. A region growing algorithm is then used to extract the complete point cloud of the damaged area, and a precise three-dimensional damage model is generated through Poisson surface reconstruction. This provides a three-dimensional geometric basis for subsequent repair area determination and adaptive layering slicing.

[0027] In some embodiments, step S20 specifically includes the following steps: calculating the maximum depth of damage based on the three-dimensional damage model; determining the repairability of the damage to be repaired based on a preset repair feasibility criterion, wherein the pre-repair feasibility criterion includes that the maximum depth of damage not exceeding a threshold proportion of the component wall thickness is repairable, otherwise it is not repairable; for damage determined to be repairable, expanding the transition zone outward based on the three-dimensional boundary of the three-dimensional damage model to obtain the repair area; the transition zone includes a constant width zone near the damage side and a gradient width zone away from the damage side, the width of the gradient width zone decreasing to zero from the inside out; and dividing the repair area into a bottom bonding sub-region, a main body filling sub-region, and a surface transition sub-region according to the damage depth distribution of the three-dimensional damage model.

[0028] The bottom bonding sub-region is sliced ​​with a first layer thickness, the main filling sub-region is sliced ​​with a second layer thickness, and the surface transition sub-region is sliced ​​with decreasing layer thicknesses; the first layer thickness is less than the second layer thickness; the first layer thickness ranges from 10μm to 20μm, and the second layer thickness ranges from 30μm to 50μm. A parallel adaptive layer slicing algorithm is used to synchronously perform slicing processing on each repair area, generating slice contour data for each layer; the exposed area area and maximum fill distance are calculated from the slice contour data, and a repair slice dataset is generated based on the slice contour data, exposed area area, and maximum fill distance.

[0029] Specifically, the system calculates the maximum depth of damage based on the 3D damage model and compares it with the component wall thickness threshold to determine repairability, avoiding ineffective repair operations on irreparable damage. For repairable damage, a transition zone containing a constant width region and a gradient width region is extended outward along its 3D boundary to determine the repair area. The gradient width region decreases from the inside to zero from the outside, achieving a smooth connection between the repair area and the original surface. Based on the damage depth distribution, the repair area is divided into a bottom bonding sub-region, a main filling sub-region, and a surface transition sub-region. A thinner layer thickness is used to ensure the metallurgical bonding strength between the bottom and the substrate, a conventional layer thickness is used to improve filling efficiency, and a decreasing layer thickness is used to achieve a smooth surface transition. Each sub-region is processed synchronously through a parallel adaptive layer slicing algorithm to generate a repair slice dataset containing slice outlines, exposure areas, and maximum filling distances, providing layered data that matches the damage depth gradient for subsequent path planning.

[0030] In some embodiments, step S30 specifically includes the following steps: mapping the deposition filling units in the repair slice dataset to cities in the coloring traveling salesman problem, mapping the repair nozzles to merchants, and mapping the process constraint types to colors; the process constraint types include overlap rate constraints, heat input uniformity constraints, scan direction constraints, and partition transition constraints; constructing a composite cost objective function for the coloring traveling salesman problem with the optimization objective of minimizing the weighted sum of total empty travel distance and heat accumulation penalty; wherein, cities of the same color can only be visited by merchants carrying the same color, each city is visited only once, and the path of each merchant forms an open-loop Hamiltonian path; A partially optimized metaheuristic decomposition framework is adopted. All deposition filling units in the current slice layer are spatially clustered by K-means clustering to obtain several clusters. The dominant process constraint type of each cluster is determined as the main color. For each cluster, the deposition filling unit corresponding to its main color is used as the core, and the deposition filling units located at the boundary in the adjacent clusters are introduced to construct an open-loop coloring traveling salesman subproblem model with a global perspective. A hybrid heuristic algorithm combining simulated annealing and variable neighborhood search is employed to solve each open-loop colored traveling salesman problem (TSP) subproblem model. The algorithm uses a dual-chromosome encoding method: the first chromosome marks the access order of deposition filling units, and the second chromosome marks the number of the repair nozzles performing deposition. In the initialization phase, a nearest-neighbor greedy strategy is used to generate initial solutions. In the neighborhood search phase, reverse ordering operators, exchange operators, and gene fragment re-insertion operators are used alternately. The solution is updated using the simulated annealing Metropolis criterion. The solutions to all open-loop colored TSP subproblems are greedily integrated to obtain an initial complete path sequence. Path repair is performed on the initial complete path sequence, checking time window constraints, modular constraints, corner constraints, and thermal accumulation constraints. Cooling waiting time is inserted for path segments that violate thermal accumulation constraints to obtain the optimal deposition path sequence.

[0031] Specifically, process constraints such as deposition filling units, repair nozzles and overlap rates, and heat input uniformity are mapped to cities, merchants, and colors in a coloring traveling salesman problem, respectively. A composite cost function is constructed with the weighted sum of empty driving distance and heat accumulation penalty as the objective. K-means clustering is used to divide large-scale deposition units into several clusters, and the dominant process constraint of each cluster is determined as the main color. An open-loop sub-problem model with a global perspective is constructed by introducing adjacent cluster boundary units with the main color unit as the core, avoiding local optimization from deviating from the global optimum. A hybrid heuristic algorithm of simulated annealing-variable neighborhood search with dual chromosome encoding is used to solve each sub-problem through neighbor search with three types of operators: nearest neighbor greedy initialization, reverse ordering, and exchange and gene fragment re-insertion, and updating with Metropolis criterion. After greedy integration, the optimal deposition path sequence is obtained by checking and correcting constraints such as time window, corner, and heat accumulation. This systematically incorporates complex process constraints into path planning, greatly reducing the proportion of empty driving paths and solution time, and solving the problem that traditional regular paths cannot adapt to irregular damage morphologies.

[0032] In some embodiments, step S40 specifically includes the following steps: constructing a multiphysics coupling model for metal-oriented energy deposition; the multiphysics coupling model for metal-oriented energy deposition includes a laser energy distribution model, a molten pool temperature field model, and a molten pool fluid dynamics model; the laser energy distribution model uses a Gaussian surface heat source model to describe the energy distribution of the laser beam on the deposition surface, as shown in the following formula: ; in, For heat flux density, Let be the absorption rate of the material to the laser, P be the laser power, R be the spot radius, and r be the distance from the center of the spot.

[0033] The molten pool temperature field model establishes a three-dimensional transient temperature field control equation based on Fourier's law of heat conduction, as shown in the following formula: ; in, Where C is the material density, Cp is the specific heat of the material, and k is the thermal conductivity of the material. The latent heat of fusion of the material, The solid fraction, For temperature; The molten pool hydrodynamic model satisfies the Navier-Stokes equations of mass and momentum conservation, and the driving forces include buoyancy driving terms and Marangoni convection terms.

[0034] Using the optimal deposition path sequence as a spatial constraint, the sequence and location information of each path segment in the optimal deposition path sequence are transformed into the spatiotemporal trajectory of the moving heat source. The spatiotemporal trajectory is loaded into the multiphysics coupling model of metal directional energy deposition through a user-defined function for numerical simulation. The numerical simulation uses fine mesh for the deposition channel and heat-affected zone, and coarse mesh for areas far from the deposition zone. The process status data of each path segment are extracted from the output of the numerical simulation. The process status data includes the maximum temperature of the molten pool, cooling rate, peak residual stress, and deposition layer thickness. Process status data is used to construct a training dataset to train an improved long short-term memory (LSTM) network. A batch normalization layer is placed between the output of the gating unit of the original LSM network and the input of the next layer to prevent gradient vanishing. A dropout layer randomly discards neurons during training at a preset dropout rate to prevent overfitting. The slice data of the model to be repaired and the optimal deposition path sequence are input into the trained improved LSM network to predict the optimal combination of process parameters for each path segment layer by layer. The optimal combination of process parameters includes laser power, scanning speed, powder feeding rate, and protective gas flow rate. If the highest temperature of the molten pool predicted by the improved LSM network exceeds the preset proportional threshold of the material flash point, a cooling waiting strategy is automatically inserted in the corresponding path segment. The cooling waiting strategy includes at least one of extending the black screen time and increasing the protective gas flow rate. Based on the optimal combination of process parameters and the cooling waiting strategy for each path segment, an optimized process parameter package is generated.

[0035] Specifically, a multiphysics coupled model for metal-oriented energy deposition is constructed, incorporating a Gaussian surface heat source laser energy distribution model, a Fourier heat conduction molten pool temperature field model, and a Navier-Stokes molten pool fluid dynamics model. Using the optimal deposition path sequence as spatial constraints, the order and position of each path segment are transformed into the spatiotemporal trajectory of the moving heat source and loaded into the model via a user-defined function for numerical simulation. Fine meshes are used for the deposition channel and heat-affected zone, while coarse meshes are used for areas far from the deposition zone to balance accuracy and efficiency. Process state data such as the highest molten pool temperature, cooling rate, peak residual stress, and deposition layer thickness are extracted from this data. This data is then used to train an improved long short-term memory network with batch normalization and discard layers to avoid gradient vanishing and overfitting problems in deep network training. After training, the model slice data to be repaired and the path sequence are input into the network to quickly predict optimized process parameters such as laser power, scanning speed, powder feeding rate, and protective gas flow rate for each path segment. A cooling waiting strategy is automatically inserted when the predicted temperature exceeds the material flash point threshold. This overcomes the shortcomings of large prediction errors in traditional empirical formulas and the time-consuming computation of pure numerical simulation, achieving synergistic optimization of process parameters and deposition paths.

[0036] In some embodiments, step S50 specifically includes the following steps: real-time acquisition of repair status data using a molten pool vision camera, an infrared thermal imager, a laser displacement sensor, and an acoustic emission sensor; wherein, the molten pool vision camera acquires molten pool morphology images at a first acquisition frequency, the infrared thermal imager acquires molten pool temperature field images at a second acquisition frequency, the laser displacement sensor acquires deposition layer thickness data at a third acquisition frequency, and the acoustic emission sensor acquires acoustic emission signals at a fourth acquisition frequency; the process status data is used as the predicted value of the extended Kalman filter, and the repair status data is used as the measured value of the extended Kalman filter.

[0037] Define the state variables and control inputs of the extended Kalman filter; the state variables include the melt pool temperature, the deposited layer thickness, and the cooling rate; the control inputs include the laser power correction and the scanning speed correction; use an improved long short-term memory network as the state transition function, and calculate the predicted value of the state variables and the prediction error covariance at the current time based on the estimated value of the state variables at the previous time step and the control input at the current time step.

[0038] Using the molten pool temperature measured by an infrared thermal imager and the deposition layer thickness measured by a laser displacement sensor as observed variables, the Kalman gain is calculated. The state variables are then corrected using the Kalman gain, observed variables, and predicted values ​​to obtain the corrected state variables, and the prediction error covariance is updated. The temperature deviation between the molten pool temperature and the highest molten pool temperature predicted by the improved long short-term memory network in the corrected state variables is calculated, as is the height deviation between the deposition layer thickness and the target deposition layer thickness in the corrected state variables. When the temperature deviation exceeds a preset temperature threshold or the height deviation exceeds a preset proportional threshold for the target deposition layer thickness, state correction is triggered. Based on the corrected state variables, the laser power and scanning speed of the corresponding path segment in the optimized process parameter package are adjusted to obtain the corrected control parameters.

[0039] The preset temperature threshold is the measurement accuracy of the infrared thermal imager, and the preset ratio threshold is 10% of the target deposition layer thickness.

[0040] Specifically, multi-source heterogeneous repair status data is constructed by synchronously acquiring molten pool morphology, temperature field, deposition layer thickness, and acoustic emission signals at different frequencies using a molten pool vision camera, an infrared thermal imager, a laser displacement sensor, and an acoustic emission sensor. The predicted process status is used as a priori estimate for the extended Kalman filter, with the real-time acquired status data as the observed values. Molten pool temperature, deposition layer thickness, and cooling rate are defined as state variables, while laser power and scanning speed corrections are used as control inputs. An improved long short-term memory network is used as the state transition function to calculate the current state prediction value and prediction error covariance. The Kalman gain is calculated using the measured molten pool temperature from the infrared thermal imager and the measured deposition layer thickness from the laser displacement sensor as observed variables, correcting the state variables and updating the covariance. When the deviation between the corrected molten pool temperature and the predicted value exceeds the measurement accuracy of the infrared thermal imager, or the deviation between the deposition layer thickness and the target thickness exceeds a preset proportional threshold, the laser power and scanning speed of the corresponding path segment are adjusted based on the corrected state variables to obtain the corrected control parameters. By dynamically fusing physical model predictions with multi-source measured data, the impact of accumulated deviations in open loops on repair quality is effectively eliminated.

[0041] In some embodiments, step S60 specifically includes the following steps: identifying the type of repair anomaly based on the molten pool morphology image, deposition layer thickness data, molten pool temperature field image, and acoustic emission signal in the repair status data; the types of repair anomalies include: lack of fusion, over-deposition, surface cracks, spatter, porosity, and excessive heat accumulation; and selecting a corresponding repair strategy from a preset set of repair strategies according to the type of repair anomaly; the set of repair strategies includes: local remelting strategy, in-situ subtractive material correction strategy, crack filling strategy, surface cleaning and re-deposition strategy, and forced cooling strategy.

[0042] When the repair anomaly type is non-fusion, a local remelting strategy is invoked. This strategy includes: increasing laser power based on the corrected control parameters, decreasing the scanning speed based on the corrected control parameters, and performing a second scan deposition on the path segment corresponding to the non-fusion. When the repair anomaly type is over-deposition, an in-situ subtractive correction strategy is invoked. This strategy includes: using the height deviation of the deposited layer thickness in the corrected control parameters as the cutting amount, invoking the accompanying subtractive tool to perform in-situ height correction on the path segment corresponding to the over-deposition, and then performing deposition based on the corrected control parameters. When the repair anomaly type is surface crack, a crack filling strategy is invoked. The strategies include: reducing the scanning speed based on the corrected control parameters, adjusting the powder feeding rate according to the corrected control parameters, and performing crack filling deposition along the crack direction; when the repair anomaly type is spatter or porosity, invoking the surface cleaning and re-deposition strategy; the surface cleaning and re-deposition strategy includes: using an online polishing tool to clean the surface of the path segment corresponding to spatter or porosity, and performing fixed-point re-deposition at the cleaned position according to the corrected control parameters; when the repair anomaly type is excessive heat accumulation, invoking the forced cooling strategy; the forced cooling strategy includes: inserting a cooling waiting time in the path segment corresponding to excessive heat accumulation, and adjusting the protective gas flow rate based on the corrected control parameters.

[0043] Specifically, the system identifies anomalies such as incomplete fusion, over-deposition, surface cracks, spatter, porosity, and excessive heat accumulation based on the molten pool morphology image, deposition layer thickness, molten pool temperature field, and acoustic emission signals in the repair status data. According to the anomaly type, the system retrieves the corresponding strategy from a preset strategy set. Each strategy uses the corrected control parameters as the execution benchmark: for incomplete fusion, the laser power is increased and the scanning speed is reduced for a second scan deposition to replenish the fusion energy; for over-deposition, the accompanying subtractive material tool is used to correct the height in situ and redeposit, based on the corrected height deviation; for surface cracks, the scanning speed is reduced and the powder feed rate is adjusted to fill the crack along its direction; for spatter or porosity, the surface is cleaned using an online polishing tool and then targeted deposition is performed; for excessive heat accumulation, a cooling waiting time is inserted and the protective gas flow is increased to accelerate heat dissipation. This binds anomaly identification with online repair strategies, driving various repair operations with corrected real-time control parameters. This avoids the problem of traditional open-loop methods where defects necessitate system shutdown and rework, significantly improving repair success rate and reliability.

[0044] The implementation principle of the path planning method for repairing surface damage of metal parts based on artificial intelligence in this application embodiment is as follows: The system uses an embedded coordinate attention mechanism, adds a minimal damage detection layer, and uses an improved YOLO detection network with an EIOU loss function to detect damage on the surface image of the metal part. The detection results are mapped to depth point cloud data, and a three-dimensional damage model is generated through region growing and Poisson surface reconstruction. Based on the three-dimensional damage model, the repair area is determined. According to the damage depth distribution, the bottom bonding area, the main filling area, and the surface transition area are adaptively sliced ​​with different layer thicknesses to generate a repair slice dataset. The repair slice dataset is mapped to a colorized traveling salesman problem. The problem is divided into sub-problems by K-means clustering and solved using a hybrid heuristic algorithm of simulated annealing-variable neighborhood search. The optimal deposition path sequence is obtained through path repair and greedy integration. A multiphysics coupled model for metal-oriented energy deposition (MED) was constructed, incorporating laser energy distribution, molten pool temperature field, and molten pool fluid dynamics. Numerical simulations were performed using the optimal deposition path sequence as spatial constraints. An improved long short-term memory (LSTM) network with batch normalization and discard layers was trained using simulation data to predict the optimal combination of process parameters for each path segment, forming an optimized process parameter package. During the repair process, repair status data was collected in real-time using multi-source sensors. An extended Kalman filter was used to fuse the predicted values ​​from the improved LSD network with the measured values. When the deviation exceeded a preset threshold, a status correction was triggered, and corrected control parameters were obtained. When a repair anomaly was detected, the corresponding repair strategy was invoked based on the corrected control parameters for online anomaly repair. This approach achieved high-precision localization and 3D reconstruction of surface damage in metal parts, efficient planning of deposition paths, real-time dynamic optimization of process parameters, and online repair of various anomalies. It significantly improved repair accuracy and efficiency, greatly reduced the proportion of empty paths, and minimized process parameter prediction errors, providing a complete technical solution for ultra-micro 3D printing repair of surface damage in metal parts.

[0045] Figure 1 This is a flowchart illustrating a path planning method for repairing surface damage of metal parts using artificial intelligence-based ultra-micro 3D printing, as shown in one embodiment. It should be understood that, although... Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows; unless explicitly stated otherwise, there is no strict order requirement for the execution of these steps, and they can be executed in other orders; and Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0046] Based on the same technical concept, referring to Figure 2 This application also provides a path planning device for repairing surface damage of metal parts using artificial intelligence-based micro 3D printing, which adopts the following technical solution: The device includes: The 3D damage module is used to perform damage detection on surface images of metal parts using an improved YOLO detection network, map the detection results to depth point cloud data, and generate a 3D damage model through region growing and Poisson surface reconstruction. The repair data module is used to determine the repair area based on the 3D damage model, perform adaptive layered slicing, and generate a repair slice dataset. The deposition path module maps the repair slice dataset to the colorized traveling salesman problem. It uses a partially optimized metaheuristic decomposition framework to decompose the subproblem through K-means clustering and solves it using a hybrid heuristic algorithm of simulated annealing-variable neighborhood search. After path repair and greedy integration, the optimal deposition path sequence is obtained. The optimization process module is used to construct a multiphysics coupling model for metal directional energy deposition. It performs numerical simulation with the optimal deposition path sequence as a spatial constraint, and uses the simulation data to train an improved long short-term memory network with added batch normalization and discard layers to predict the optimal combination of process parameters for each path segment and obtain an optimized process parameter package. The parameter correction module is used to collect repair status data in real time. It uses extended Kalman filtering to fuse the predicted values ​​of the improved long short-term memory network with the measured values. When the deviation exceeds the preset threshold, the status correction is triggered to obtain the correction control parameters. The repair control module is used to call the corresponding repair strategy to perform abnormal repair when a repair anomaly is detected, based on the correction control parameters.

[0047] In some embodiments, the damage 3D module is specifically used to acquire images of the surface of a metal part using an industrial camera under multi-angle and multi-light conditions. The surface image of the metal part is input into the improved YOLO detection network to obtain a two-dimensional damage bounding box and damage classification results. The damage classification results include at least one of cracks, wear pits, pitting, scratches, spalling and oxidation corrosion pits. The two-dimensional bounding box of the damage is mapped to the depth point cloud data obtained by structured light or laser line scanning to determine the three-dimensional boundary of the damaged area; Based on the three-dimensional boundary of the damaged region, a region growing algorithm is used to extract the complete point cloud of the damaged region from the depth point cloud data. The Poisson surface reconstruction algorithm is used to reconstruct the surface of the complete point cloud and generate a damaged 3D model.

[0048] In some embodiments, the repair data module is specifically used to calculate the maximum depth of damage based on the three-dimensional damage model, and to determine the repairability of the damage to be repaired based on a preset repair feasibility criterion. The pre-repair feasibility criterion includes that the maximum depth of damage does not exceed a threshold proportion of the component wall thickness, which is repairable; otherwise, it is not repairable. For damage that is determined to be repairable, the transition zone is expanded outward based on the three-dimensional boundary of the damage three-dimensional model to obtain the repair area; the transition zone includes a constant width zone near the damage side and a gradient width zone away from the damage side, and the width of the gradient width zone decreases to zero from the inside to the outside. Based on the damage depth distribution of the three-dimensional damage model, the repair area is divided into a bottom bonding sub-region, a main body filling sub-region, and a surface transition sub-region. The bottom bonding sub-region is sliced ​​with a first layer thickness, the main filling sub-region is sliced ​​with a second layer thickness, and the surface transition sub-region is sliced ​​with a decreasing layer thickness; the first layer thickness is less than the second layer thickness. A parallel adaptive layered slicing algorithm is used to synchronously perform slicing processing on each repair area to generate slice contour data for each layer. The exposed area and maximum fill distance are calculated from the slice contour data. Based on the slice contour data, exposed area and maximum fill distance, a repair slice dataset is generated.

[0049] In some embodiments, the deposition path module is specifically used to map the deposition fill cells in the repair slice dataset to cities in a colorized traveling salesman problem, to repair nozzles to merchants, and to process constraint types to colors; the process constraint types include overlap ratio constraints, thermal input uniformity constraints, scan direction constraints, and partition transition constraints. With the goal of minimizing the weighted sum of total empty travel distance and heat accumulation penalty, a composite cost objective function for the colored traveling salesman problem is constructed. In this function, cities of the same color can only be visited by merchants carrying the same color, each city is visited exactly once, and each merchant's path forms an open-loop Hamiltonian path. A partially optimized metaheuristic decomposition framework is adopted, and K-means clustering is used to spatially cluster all deposition filling units of the current slice layer to obtain several clusters; the dominant process constraint type of each cluster is determined as the main color; For each cluster, the depositional filling unit corresponding to its main color is used as the core, and the depositional filling units located at the boundary of the adjacent clusters are introduced to construct an open-loop coloring traveling salesman subproblem model with a global perspective. A hybrid heuristic algorithm combining simulated annealing and variable neighborhood search is used to solve each open-loop colored traveling salesman problem model. The algorithm employs a dual-chromosome encoding method: the first chromosome marks the access order of the deposition filling units, and the second chromosome marks the number of the repair nozzles performing the deposition. In the initialization phase, a nearest-neighbor greedy strategy is used to generate the initial solution. In the neighborhood search phase, the reverse order operator, the exchange operator, and the gene fragment re-insertion operator are used alternately. The solution is updated using the simulated annealing Metropolis criterion. Greedy integration of solutions to all open-loop colored traveling salesman subproblem models yields an initial complete path sequence; The initial complete path sequence is repaired by checking time window constraints, modular constraints, corner constraints and thermal accumulation constraints. Cooling waiting time is inserted for path segments that violate thermal accumulation constraints to obtain the optimal deposition path sequence.

[0050] In some embodiments, the process optimization module is specifically used to construct a multiphysics coupling model for metal directional energy deposition; the multiphysics coupling model for metal directional energy deposition includes a laser energy distribution model, a molten pool temperature field model, and a molten pool fluid dynamics model. Among them, the laser energy distribution model uses the Gaussian surface heat source model to describe the energy distribution of the laser beam on the deposition surface; the molten pool temperature field model establishes the three-dimensional transient temperature field control equation based on Fourier's heat conduction law; the molten pool fluid dynamics model satisfies the Navier-Stokes equations of mass conservation and momentum conservation, and the driving force includes buoyancy driving term and Marangoni convection term. Using the optimal deposition path sequence as a spatial constraint, the sequence and location information of each path segment in the optimal deposition path sequence are transformed into the spatiotemporal trajectory of the moving heat source. The spatiotemporal trajectory is loaded into the multiphysics coupling model of metal directional energy deposition through a user-defined function for numerical simulation. The numerical simulation uses fine mesh for the deposition channel and thermally affected zone, and coarse mesh for areas far from the deposition zone. Extract process status data for each path segment from the output of the numerical simulation; process status data includes the maximum temperature of the molten pool, cooling rate, peak residual stress, and deposition layer thickness; Process status data is used to construct a training dataset to train an improved long short-term memory network. Batch normalization layers are placed between the output of the gating unit of the original long short-term memory network and the input of the next layer to prevent gradient vanishing. Dropout layers randomly drop neurons during training at a preset dropout rate to prevent overfitting. The slice data of the model to be repaired and the optimal deposition path sequence are input into the trained improved long short-term memory network to predict the optimal combination of process parameters for each path segment layer by layer; the optimal combination of process parameters includes laser power, scanning speed, powder feeding rate and protective gas flow rate. If the maximum temperature of the molten pool predicted by the improved long short-term memory network exceeds the preset proportional threshold of the material's flash point, a cooling waiting strategy will be automatically inserted in the corresponding path segment; the cooling waiting strategy includes at least one of extending the black screen time and increasing the protective gas flow rate. Based on the optimal combination of process parameters and cooling waiting strategies for each path segment, an optimized process parameter package is generated.

[0051] In some embodiments, the parameter correction module is specifically used to collect repair status data in real time through a molten pool vision camera, an infrared thermal imager, a laser displacement sensor, and an acoustic emission sensor; wherein, the molten pool vision camera collects molten pool morphology images at a first acquisition frequency, the infrared thermal imager collects molten pool temperature field images at a second acquisition frequency, the laser displacement sensor collects deposition layer thickness data at a third acquisition frequency, and the acoustic emission sensor collects acoustic emission signals at a fourth acquisition frequency. Process status data is used as the predicted value of extended Kalman filter, and repair status data is used as the measured value of extended Kalman filter. Define the state variables and control inputs of the extended Kalman filter; the state variables include the melt pool temperature, the deposited layer thickness, and the cooling rate; the control inputs include the laser power correction and the scanning speed correction. Using an improved long short-term memory network as the state transition function, the predicted value of the state variable at the current time and the prediction error covariance are calculated based on the estimated value of the state variable at the previous time step and the control input at the current time step. The molten pool temperature measured by an infrared thermal imager and the deposition layer thickness measured by a laser displacement sensor are used as observation variables. The Kalman gain is calculated, and the state variables are corrected using the Kalman gain, the observation variables, and the predicted values ​​to obtain the corrected state variables. The prediction error covariance is then updated. Calculate the temperature deviation between the molten pool temperature and the maximum molten pool temperature predicted by the improved long short-term memory network in the corrected state variables, and calculate the height deviation between the deposition layer thickness and the target deposition layer thickness in the corrected state variables. When the temperature deviation exceeds the preset temperature threshold or the height deviation exceeds the preset proportional threshold of the target deposition layer thickness, state correction is triggered; based on the corrected state variables, the laser power and scanning speed of the corresponding path segment in the optimized process parameter package are adjusted to obtain the corrected control parameters.

[0052] In some embodiments, the repair control module is specifically used to identify repair anomaly types based on the molten pool morphology image, deposition layer thickness data, molten pool temperature field image, and acoustic emission signal in the repair status data; repair anomaly types include: lack of fusion, excessive deposition, surface cracks, spatter, porosity, and excessive heat accumulation. Based on the type of anomaly to be repaired, a corresponding repair strategy is selected from a set of preset repair strategies. The set of repair strategies includes: local remelting strategy, in-situ subtractive material correction strategy, crack filling strategy, surface cleaning and re-deposition strategy, and forced cooling strategy. When the repair anomaly type is non-fusion, the local remelting strategy is invoked. The local remelting strategy includes: increasing the laser power based on the modified control parameters, decreasing the scanning speed based on the modified control parameters, and performing secondary scanning deposition on the path segment corresponding to the non-fusion. When the repair anomaly type is over-accumulation, the in-situ subtractive correction strategy is invoked. The in-situ subtractive correction strategy includes: using the height deviation of the deposition layer thickness in the correction control parameters as the cutting amount, invoking the accompanying subtractive tool to perform in-situ height correction on the path segment corresponding to over-accumulation, and then performing deposition with the correction control parameters. When repairing anomalies of surface cracks, a crack filling strategy is invoked. The crack filling strategy includes: reducing the scanning speed based on the modified control parameters, adjusting the powder feeding rate based on the modified control parameters, and performing crack filling deposition along the crack direction. When repairing anomalies of splash or porosity, the surface cleaning and re-deposition strategy is invoked. The surface cleaning and re-deposition strategy includes: invoking an online polishing tool to clean the path segment corresponding to the splash or porosity, and then correcting the control parameters at the cleaned position for targeted re-deposition. When the repair anomaly type is thermal accumulation exceeding the limit, the forced cooling strategy is invoked. The forced cooling strategy includes: inserting a cooling waiting time in the path segment corresponding to the thermal accumulation exceeding the limit, and adjusting the protective gas flow rate based on the corrected control parameters.

[0053] This application also discloses a control device.

[0054] Specifically, the control device includes a memory and a processor. The memory stores a computer program that can be loaded by the processor and executed to perform the aforementioned path planning method for repairing surface damage of metal parts using artificial intelligence-based micro 3D printing.

[0055] This application also discloses a computer-readable storage medium.

[0056] Specifically, the computer-readable storage medium stores a computer program that can be loaded by a processor and executed, such as the aforementioned path planning method for repairing surface damage of metal parts using artificial intelligence-based micro 3D printing. The computer-readable storage medium includes, for example, various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0057] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A path planning method for repairing surface damage of metal parts using artificial intelligence-based micro-3D printing, characterized in that, include: An improved YOLO detection network is used to perform damage detection on the surface images of metal parts. The detection results are mapped to depth point cloud data, and a three-dimensional damage model is generated through region growing and Poisson surface reconstruction. Based on the three-dimensional damage model, the repair area is determined, adaptive layered slicing is performed, and a repair slice dataset is generated. The repaired slice dataset is mapped to the Colored Traveling Salesman Problem. A partially optimized metaheuristic decomposition framework is used to decompose the subproblem through K-means clustering, and a simulated annealing-variable neighborhood search hybrid heuristic algorithm is used to solve it. After path repair and greedy integration, the optimal deposition path sequence is obtained. A multiphysics coupling model for metal directional energy deposition was constructed, and numerical simulation was performed with the optimal deposition path sequence as spatial constraints. An improved long short-term memory network with added batch normalization and discard layers was trained using the simulation data to predict the optimal combination of process parameters for each path segment and obtain the optimized process parameter package. Real-time acquisition of repair status data; use extended Kalman filtering to fuse the predicted values ​​of the improved long short-term memory network with the measured values; when the deviation exceeds a preset threshold, trigger status correction and obtain correction control parameters. When a repair anomaly is detected, the corresponding repair strategy is invoked based on the correction control parameters to perform the anomaly repair.

2. The path planning method for repairing surface damage of metal parts using artificial intelligence-based ultra-micro 3D printing, as described in claim 1, is characterized in that... The method involves using an improved YOLO detection network to perform damage detection on surface images of metal components, mapping the detection results to depth point cloud data, and generating a 3D damage model through region growing and Poisson surface reconstruction, including: The improved YOLO detection network consists of a backbone network, a neck network, and a detection head. A coordinate attention mechanism module is inserted before the fast spatial pyramid pooling module in the backbone network. This module decomposes channel attention into one-dimensional feature encoding in the horizontal direction and one-dimensional feature encoding in the vertical direction, and aggregates features along the two spatial directions to generate an attention weight matrix. A high-pixel minimal damage detection layer is added to the neck network, expanding the detection layer to four levels. The improved YOLO detection network uses an EIOU loss function based on four geometric factors: overlapping region, center point distance, width difference, and height difference, for bounding box regression. Images of metal parts are captured using industrial cameras under multiple angles and lighting conditions. The surface image of the metal part is input into the improved YOLO detection network to obtain a two-dimensional bounding box of damage and damage classification results. The damage classification results include at least one of cracks, wear pits, pitting, scratches, spalling and oxidation corrosion pits. The two-dimensional bounding box of the damage is mapped to the depth point cloud data obtained by structured light or laser line scanning to determine the three-dimensional boundary of the damaged area; Based on the three-dimensional boundary of the damaged region, a region growing algorithm is used to extract the complete point cloud of the damaged region from the depth point cloud data. The Poisson surface reconstruction algorithm is used to reconstruct the surface of the complete point cloud to generate the damaged 3D model.

3. The path planning method for repairing surface damage of metal parts using artificial intelligence-based ultra-micro 3D printing, as described in claim 2, is characterized in that... The process of determining the repair area based on the three-dimensional damage model, performing adaptive layered slicing, and generating a repair slice dataset includes: The maximum depth of the damage is calculated based on the three-dimensional damage model. The repairability of the damage to be repaired is judged based on the preset repair feasibility criteria. The preset repair feasibility criteria include that the maximum depth of the damage does not exceed a threshold proportion of the component wall thickness, which is repairable; otherwise, it is not repairable. For damage that is determined to be repairable, a transition zone is expanded outward based on the three-dimensional boundary of the damage three-dimensional model to obtain a repair area; the transition zone includes a constant width zone near the damage side and a gradient width zone away from the damage side, and the width of the gradient width zone decreases to zero from the inside to the outside. Based on the damage depth distribution of the three-dimensional damage model, the repair area is divided into a bottom bonding sub-region, a main body filling sub-region, and a surface transition sub-region. The bottom bonding sub-region is sliced ​​with a first layer thickness, the main filling sub-region is sliced ​​with a second layer thickness, and the surface transition sub-region is sliced ​​with a decreasing layer thickness; the first layer thickness is less than the second layer thickness. A parallel adaptive layered slicing algorithm is used to synchronously perform slicing processing on each of the repair areas to generate slice contour data for each layer; the exposed area area and the maximum fill distance are calculated on the slice contour data, and the repair slice dataset is generated based on the slice contour data, the exposed area area and the maximum fill distance.

4. The path planning method for repairing surface damage of metal parts using artificial intelligence-based ultra-micro 3D printing, as described in claim 1, is characterized in that... The process involves mapping the repaired slice dataset to a colorized traveling salesman problem, employing a partially optimized metaheuristic decomposition framework to decompose subproblems using K-means clustering, and solving them using a simulated annealing-variable neighborhood search hybrid heuristic algorithm. After path repair and greedy integration, the optimal deposition path sequence is obtained, including: The deposition filling units in the repair slice dataset are mapped to cities in the Colored Traveling Salesman Problem, the repair nozzles are mapped to merchants, and the process constraint types are mapped to colors; the process constraint types include overlap ratio constraints, thermal input uniformity constraints, scan direction constraints, and partition transition constraints. With the goal of minimizing the weighted sum of total empty travel distance and heat accumulation penalty, a composite cost objective function for the colored traveling salesman problem is constructed. In this function, cities of the same color can only be visited by merchants carrying the same color, each city is visited exactly once, and each merchant's path forms an open-loop Hamiltonian path. A partially optimized metaheuristic decomposition framework is adopted, and K-means clustering is used to spatially cluster all the deposition filling units of the current slice layer to obtain several clusters; the dominant process constraint type of each cluster is determined as the main color; For each cluster, the deposition filling unit corresponding to its primary color is used as the core, and the deposition filling units located at the boundary of the adjacent clusters are introduced to construct an open-loop coloring traveling salesman problem model with a global perspective. A hybrid heuristic algorithm of simulated annealing and variable neighborhood search is used to solve each of the open-loop coloring traveling salesman problem models. The simulated annealing-variable neighborhood search hybrid heuristic algorithm employs a dual-chromosome encoding method, where the first chromosome marks the access order of the deposition filling units, and the second chromosome marks the number of the repair nozzle performing the deposition. In the initialization phase, a nearest-neighbor greedy strategy is used to generate an initial solution. In the neighborhood search phase, the inversion operator, the exchange operator, and the gene fragment re-insertion operator are used alternately. The solution is updated using the simulated annealing Metropolis criterion. Greedy integration of solutions to all the open-loop colored traveling salesman subproblem models yields an initial complete path sequence; The initial complete path sequence is repaired by checking time window constraints, modular constraints, corner constraints and thermal accumulation constraints. Cooling waiting time is inserted for path segments that violate the thermal accumulation constraints to obtain the optimal deposition path sequence.

5. The path planning method for repairing surface damage of metal parts using artificial intelligence-based ultra-micro 3D printing, as described in claim 4, is characterized in that... The construction of a multiphysics coupled model for metal-oriented energy deposition involves numerical simulation using the optimal deposition path sequence as a spatial constraint. An improved long short-term memory network with added batch normalization and discard layers is trained using the simulation data to predict the optimal combination of process parameters for each path segment, thereby obtaining an optimized process parameter package, including: A multiphysics coupling model for metal directional energy deposition is constructed; the multiphysics coupling model for metal directional energy deposition includes a laser energy distribution model, a molten pool temperature field model, and a molten pool fluid dynamics model. The laser energy distribution model uses a Gaussian surface heat source model to describe the energy distribution of the laser beam on the deposition surface; the molten pool temperature field model establishes a three-dimensional transient temperature field control equation based on Fourier's heat conduction law; the molten pool fluid dynamics model satisfies the Navier-Stokes equations of mass conservation and momentum conservation, and the driving force includes buoyancy driving terms and Marangoni convection terms. Using the optimal deposition path sequence as a spatial constraint, the sequence and position information of each path segment in the optimal deposition path sequence are transformed into the spatiotemporal trajectory of the moving heat source; the spatiotemporal trajectory is loaded into the metal directional energy deposition multiphysics coupling model through a user-defined function for numerical simulation; the numerical simulation uses fine mesh for the deposition channel and thermally affected zone, and coarse mesh for areas far from the deposition zone; The process status data for each path segment is extracted from the output of the numerical simulation; the process status data includes the maximum temperature of the molten pool, the cooling rate, the peak residual stress, and the thickness of the deposited layer; The process state data is used to construct a training dataset to train an improved long short-term memory network. A batch normalization layer is placed between the output of the gating unit of the original long short-term memory network and the input of the next layer to prevent gradient vanishing. A dropout layer randomly drops neurons during training at a preset dropout rate to prevent overfitting. The slice data of the model to be repaired and the optimal deposition path sequence are input into the trained improved long short-term memory network to predict the optimal process parameter combination for each path segment layer by layer; the optimal process parameter combination includes laser power, scanning speed, powder feeding rate and protective gas flow rate. If the maximum temperature of the molten pool predicted by the improved long short-term memory network exceeds a preset proportional threshold of the material's flash point, a cooling waiting strategy is automatically inserted in the corresponding path segment; the cooling waiting strategy includes at least one of extending the black screen time and increasing the protective gas flow rate. Based on the optimal combination of process parameters for each path segment and the cooling waiting strategy, the optimized process parameter package is generated.

6. The path planning method for repairing surface damage of metal parts using artificial intelligence-based ultra-micro 3D printing, as described in claim 5, is characterized in that... The real-time acquisition and repair status data uses an extended Kalman filter to fuse the predicted values ​​of the improved long short-term memory network with the measured values. When the deviation exceeds a preset threshold, status correction is triggered, and correction control parameters are obtained, including: The repair status data is collected in real time using a molten pool vision camera, an infrared thermal imager, a laser displacement sensor, and an acoustic emission sensor. The molten pool vision camera collects molten pool morphology images at a first acquisition frequency, the infrared thermal imager collects molten pool temperature field images at a second acquisition frequency, the laser displacement sensor collects deposition layer thickness data at a third acquisition frequency, and the acoustic emission sensor collects acoustic emission signals at a fourth acquisition frequency. The process status data is used as the predicted value of the extended Kalman filter, and the repair status data is used as the measured value of the extended Kalman filter. Define the state variables and control inputs for the extended Kalman filter; the state variables include the melt pool temperature, the deposited layer thickness, and the cooling rate; the control inputs include the laser power correction and the scanning speed correction. Using the improved long short-term memory network as the state transition function, the predicted value of the state variable at the current time and the prediction error covariance are calculated based on the estimated value of the state variable at the previous time and the control input at the current time. Using the molten pool temperature measured by an infrared thermal imager and the deposition layer thickness measured by a laser displacement sensor as observation variables, the Kalman gain is calculated. The state variable is then corrected using the Kalman gain, the observation variables, and the predicted value to obtain the corrected state variable, and the prediction error covariance is updated. Calculate the temperature deviation between the corrected state variables and the maximum temperature of the melt pool predicted by the improved long short-term memory network, and calculate the height deviation between the corrected state variables and the target deposit thickness. When the temperature deviation exceeds a preset temperature threshold or the height deviation exceeds a preset proportional threshold for the target deposition layer thickness, a state correction is triggered. Based on the corrected state variable, the laser power and scanning speed of the corresponding path segment in the optimized process parameter package are adjusted to obtain the corrected control parameters.

7. The path planning method for repairing surface damage of metal parts using artificial intelligence-based ultra-micro 3D printing according to claim 6, characterized in that, When a repair anomaly is detected, the corresponding repair strategy is invoked based on the corrected control parameters to perform anomaly repair, including: Based on the molten pool morphology image, deposition layer thickness data, molten pool temperature field image, and acoustic emission signal in the repair status data, the repair anomaly type is identified; the repair anomaly types include: lack of fusion, excessive deposition, surface cracks, spatter, porosity, and excessive heat accumulation. Based on the type of repair anomaly, the corresponding repair strategy is selected from a preset set of repair strategies; the set of repair strategies includes: local remelting strategy, in-situ subtractive material correction strategy, crack filling strategy, surface cleaning and re-deposition strategy, and forced cooling strategy; When the repair anomaly type is non-fusion, the local remelting strategy is invoked; the local remelting strategy includes: increasing the laser power based on the correction control parameters, decreasing the scanning speed based on the correction control parameters, and performing secondary scanning deposition on the path segment corresponding to non-fusion. When the repair anomaly type is over-deposition, the in-situ subtractive correction strategy is invoked; the in-situ subtractive correction strategy includes: using the height deviation of the deposition layer thickness in the correction control parameters as the cutting amount, invoking the accompanying subtractive tool to perform in-situ height correction on the path segment corresponding to over-deposition, and then performing deposition using the correction control parameters; When the repair anomaly type is a surface crack, a crack filling strategy is invoked; the crack filling strategy includes: reducing the scanning speed based on the correction control parameters, adjusting the powder feeding rate with the correction control parameters, and performing crack filling deposition along the crack direction; When the repair anomaly type is splash or pore, the surface cleaning and re-deposition strategy is invoked; the surface cleaning and re-deposition strategy includes: invoking an online polishing tool to clean the path segment corresponding to the splash or pore, and performing fixed-point re-deposition at the cleaned position using the correction control parameters; When the repair anomaly type is thermal accumulation exceeding the limit, a forced cooling strategy is invoked; the forced cooling strategy includes: inserting a cooling waiting time into the path segment corresponding to the thermal accumulation exceeding the limit, and adjusting the protective gas flow rate based on the correction control parameters.

8. A path planning device for repairing surface damage of metal parts using artificial intelligence-based micro-3D printing, characterized in that, The device includes: The 3D damage module is used to perform damage detection on surface images of metal parts using an improved YOLO detection network, map the detection results to depth point cloud data, and generate a 3D damage model through region growing and Poisson surface reconstruction. The repair data module is used to determine the repair area based on the damage 3D model, perform adaptive layered slicing, and generate a repair slice dataset. The deposition path module is used to map the repair slice dataset into a coloring traveling salesman problem. It adopts a partially optimized metaheuristic decomposition framework to decompose the subproblem through K-means clustering and solves it using a simulated annealing-variable neighborhood search hybrid heuristic algorithm. After path repair and greedy integration, the optimal deposition path sequence is obtained. The optimization process module is used to construct a multiphysics coupling model for metal directional energy deposition, perform numerical simulation with the optimal deposition path sequence as a spatial constraint, train an improved long short-term memory network with added batch normalization and discard layers using simulation data, predict the optimal combination of process parameters for each path segment, and obtain an optimized process parameter package. The parameter correction module is used to collect repair status data in real time. It uses extended Kalman filtering to fuse the predicted value and measured value of the improved long short-term memory network. When the deviation exceeds the preset threshold, the status correction is triggered to obtain the correction control parameters. The repair control module is used to call the corresponding repair strategy to perform abnormal repair based on the correction control parameters when a repair anomaly is detected.

9. A control device, characterized in that, The device includes: A memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and executed as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer program is stored that can be loaded by a processor and executed as described in any one of claims 1 to 7.

Images