Electromagnetic-assisted nickel-based alloy laser melting deposition crack inhibition method and electromagnetic-assisted nickel-based alloy laser melting deposition crack inhibition system

Abstract

The invention provides an electromagnetic-assisted nickel-based alloy laser melting deposition crack inhibition method and system, relates to the technical field of automatic control, constructs an intelligent control system, carries out deep fusion and real-time prediction on multi-source heterogeneous sensing data, generates a prediction result, sets a reward function and an action-value network, and improves the performance of the nickel-based alloy laser melting deposition crack inhibition system. Solving to obtain an optimization control strategy and analyzing and outputting the optimization control strategy; and constructing a three-dimensional electromagnetic field distribution matrix. In combination with real-time feedback of the shape of the molten pool, the molten pool is stimulated to flow directionally and oscillate non-linearly according to the resonance effect, oscillation parameters are obtained, thermodynamic behaviors corresponding to the molten pool are predicted according to the oscillation parameters, and a thermodynamic prediction result is fed back to the central control unit; the method comprises the following steps: detecting the surface appearance of a part, extracting defect characteristic parameter vectors, inverting defect influence factors, generating a dynamic regulation and control curve, dispersing the dynamic regulation and control curve into multi-dimensional control instruction vectors, updating a preset regulation and control model, outputting optimized regulation and control model parameters, and carrying out crack suppression operation again.

Description

Technical Field

[0001] The present invention relates to the field of automatic control technology, and in particular to an electromagnetically assisted nickel-based alloy laser melting deposition crack suppression method and system. Background Art

[0002] Nickel-based high-temperature alloys are widely used in the manufacture of high-temperature components such as aircraft engine turbine blades and combustion chambers due to their excellent high-temperature strength, oxidation resistance and thermal corrosion resistance. However, due to the poor thermal physical properties of nickel-based alloys, crack defects are easily generated during the laser melting deposition forming process, which seriously affects the mechanical properties and service life of the workpiece. Crack defects are mainly caused by the high temperature brittleness of the material and the heat effect of welding, and manifest as intergranular and transgranular cracks, which are difficult to eliminate by simply optimizing the process parameters; Therefore, a solution is urgently needed to solve the problems existing in the prior art. Summary of the invention

[0003] The embodiments of the present invention provide an electromagnetically assisted nickel-based alloy laser melting deposition crack suppression method and system, which can at least solve some of the problems existing in the prior art.

[0004] A first aspect of an embodiment of the present invention provides an electromagnetically assisted nickel-based alloy laser melting deposition crack suppression method, comprising: Construct an intelligent control system corresponding to electromagnetic field-assisted laser melting deposition, deeply fuse the pre-collected multi-source heterogeneous sensor data according to the multi-scale convolutional neural network in the intelligent control system, extract the high-dimensional feature vector corresponding to the dynamic evolution of the molten pool, combine the long short-term memory network to make real-time predictions on the molten pool morphology, temperature distribution and stress distribution, generate prediction results and add the prediction results as the molten pool state vector to the central control unit, the central control unit sets a reward function and an action-value network with crack tendency as a constraint according to the deep reinforcement learning algorithm, solves the optimized control strategy and parses it into execution instructions output to the laser and the pre-set electromagnetic coil array; A three-dimensional electromagnetic field distribution matrix is ​​constructed based on the multi-layer electromagnetic coils in the electromagnetic coil array, orthogonal electromagnetic coils are set before and after the laser scanning direction to generate a first electromagnetic field vector, and in combination with real-time feedback of the molten pool morphology, an electromagnetic field parameter vector is dynamically generated through a generative adversarial network and parsed into a multi-dimensional control signal, and the multi-dimensional control signal is sent to the electromagnetic coil array, and in combination with the induced current, the directional flow and nonlinear oscillation of the molten pool are stimulated according to the resonance effect, the oscillation parameters are obtained, and the corresponding thermodynamic behavior of the molten pool is predicted according to the oscillation parameters, and the thermodynamic prediction results are fed back to the central control unit to optimize and produce the laser parameters and the electromagnetic field; The surface morphology of parts is detected in real time by visual sensors, and the defect characteristic parameter vector is extracted and added to the pre-trained physical drive model. The defect influencing factors are inverted in combination with the digital twin model to obtain the defect cause probability distribution and the defect risk level is obtained in combination with the dynamic evolution trend prediction. The defect cause probability distribution, the defect characteristic parameter vector and the defect risk level are used as input to generate a dynamic control curve and discretize it into a multidimensional control instruction vector. The electromagnetic field parameters and laser parameters are updated based on the multidimensional control instruction vector and a pre-set high-speed feedback channel. Through the meta-learning algorithm, the initial control model parameters are generated in combination with the manufacturing task and the material type, and the pre-set control model is updated according to the dynamic control curve. The optimized control model parameters are output and the crack suppression operation is performed again.

[0005] In an optional embodiment, An intelligent control system corresponding to electromagnetic field-assisted laser melting deposition is constructed. The multi-source heterogeneous sensor data collected in advance are deeply integrated according to the multi-scale convolutional neural network in the intelligent control system, and the high-dimensional feature vector corresponding to the dynamic evolution of the molten pool is extracted. The molten pool morphology, temperature distribution and stress distribution are predicted in real time in combination with the long short-term memory network, and the prediction results are generated and added to the central control unit as the molten pool state vector. The central control unit sets a reward function and an action-value network constrained by crack tendency according to the deep reinforcement learning algorithm, solves the optimized control strategy and parses it into execution instructions output to the laser and the pre-set electromagnetic coil array, including: The multi-source heterogeneous data corresponding to the electromagnetic field assisted laser melting deposition process is collected by pre-set sensors, wherein the multi-source heterogeneous data includes temperature field data, molten pool morphology image and material data, and an intelligent control system is constructed. For the temperature field data, the multi-receptive field convolution kernel corresponding to the multi-scale convolutional neural network in the intelligent control system is divided into multiple regions of interest. For each region of interest, the multi-receptive field convolution kernel is used to extract features to obtain a molten pool temperature distribution feature matrix. The molten pool morphology image is divided into grids of different scales by a spatial pyramid pooling algorithm, and the image features in the grid are extracted by a multi-scale convolution kernel to obtain a multi-scale feature vector of the molten pool morphology. The multi-scale feature vector of the molten pool morphology is deeply fused in combination with a cross-modal attention mechanism to obtain the high-dimensional feature vector. Based on the high-dimensional feature vector, the spatial features corresponding to the molten pool morphology are sequentially modeled through the convolutional long short-term memory network to obtain the spatiotemporal evolution law of the molten pool morphology, the molten pool temperature distribution is encoded and decoded through the encoding and decoding long short-term memory network to generate a temperature distribution map at a future moment, the stress field is weightedly modeled through the attention long and short time mechanism to obtain the stress distribution, and the spatiotemporal evolution law, temperature distribution map and stress distribution are combined to obtain the prediction result; The prediction result is added to the central control unit as a molten pool state vector. The central control unit constructs a molten pool thermal-fluid coupling model according to the heat conduction equation and the fluid dynamics equation, generates an initial control strategy and uses it as prior knowledge. The deep reinforcement learning algorithm is combined to perform data-driven strategy optimization. The crack tendency is used as a constraint condition, a reward function is constructed and an action-value network is generated. The reward function is solved to obtain an optimized control strategy, and the optimized control strategy is added to a pre-set laser and electromagnetic coil array.

[0006] In an optional embodiment, Combined with the deep reinforcement learning algorithm for data-driven strategy optimization, the crack tendency is used as a constraint condition, and the reward function is constructed as shown in the following formula: ; in, R Represents the reward value, l represents the crack tendency proportionality coefficient, e represents the strain at the current molten pool, e c represents the critical strain, β represents the temperature gradient proportionality coefficient, N Indicates the number of points in the molten pool, ▽ T i Indicates i The temperature gradient at a point, c represents the molten pool shape proportional coefficient, l represents the actual length of the molten pool, l 0 represents the target length of the molten pool, w Indicates the actual width of the molten pool, w 0 represents the target width of the molten pool, s Represents the shape error tolerance.

[0007] In an optional embodiment, A three-dimensional electromagnetic field distribution matrix is ​​constructed based on the multi-layer electromagnetic coils in the electromagnetic coil array, orthogonal electromagnetic coils are arranged before and after the laser scanning direction to generate a first electromagnetic field vector, and an electromagnetic field parameter vector is dynamically generated by a generative adversarial network in combination with real-time feedback of the molten pool morphology and parsed into a multi-dimensional control signal, and the multi-dimensional control signal is sent to the electromagnetic coil array, including: Based on the multi-layer electromagnetic coils in the electromagnetic coil array, a three-dimensional electromagnetic field distribution matrix is ​​constructed by three-dimensional modeling software, orthogonal coils are arranged in the front and rear directions of the electromagnetic coil array, and the electromagnetic field vector is adjusted by adjusting the current phase and amplitude to obtain the first electromagnetic field vector; A molten pool morphology image is acquired by a pre-set high-speed camera, the acquired molten pool morphology image is pre-processed, the molten pool edge contour is extracted, the molten pool area is calculated by combining the semi-major axis and the semi-minor axis corresponding to the molten pool obtained by fitting the elliptic curve, the depth of the molten pool is estimated by combining the grayscale change, a neural network model is constructed and the current electromagnetic field parameters are used as input, and the molten pool characteristic parameters composed of the molten pool area, the molten pool depth and the molten pool edge contour are used as output, a mapping relationship is fitted and a feedback signal is generated; The generator in the generative adversarial network receives random noise and the feedback signal, dynamically generates an electromagnetic field parameter vector in combination with the first electromagnetic field vector, predicts the corresponding molten pool state and compares it with the feedback signal, determines the adversarial loss and optimizes the parameters of the generator, repeats generation and discrimination, uses the electromagnetic field parameter vector as the action space of reinforcement learning, uses the deviation between the molten pool characteristic parameters and the target parameters as the reward function, updates the weights of the generative adversarial network in combination with the gradient strategy and generates the optimal electromagnetic field parameters, and parses the optimal electromagnetic field parameters into multidimensional control signals and sends them to the electromagnetic coil array.

[0008] In an optional embodiment, Combined with the induced current, the directional flow and nonlinear oscillation of the molten pool are excited according to the resonance effect, the oscillation parameters are obtained and the thermodynamic behavior corresponding to the molten pool is predicted according to the oscillation parameters, the thermodynamic prediction results are fed back to the central control unit, and the laser parameters and electromagnetic fields are optimized and produced, including: Based on the multi-dimensional control signal and the induced current in the electromagnetic field, the volume electromagnetic force and the skin depth corresponding to the electromagnetic field frequency are calculated, the natural oscillation frequency of the molten pool is calculated based on the molten pool morphology, the flow and heat transfer behavior of the molten pool is described according to the fluid-solid coupling model, the boundary conditions are set and the control equations are discretized in combination with the preferential volume method, the physical quantities are stored in a staggered grid and the second-order upwind format is used for the flow direction, the diffusion term is second-order differentiated, the molten pool oscillation amplitude and oscillation frequency characteristics are solved, the molten pool oscillation amplitude and the oscillation frequency characteristics are combined into oscillation parameters and output; Based on the oscillation parameters, an electromagnetic induced oscillation experiment is carried out according to the induced current and the resonance effect, and a mapping model between the oscillation parameters and the thermodynamic parameters of the molten pool is constructed. The pre-set nonlinear regression model is trained in combination with the machine learning algorithm to obtain a prediction result, wherein the prediction result is the thermodynamic behavior corresponding to the molten pool, and the prediction result is fed back to the central control unit. By comparing the prediction result with the pre-set target result, an error value is generated, and the laser parameters and the electromagnetic field are optimized according to the error value, and production is carried out according to the optimized parameters, wherein the thermodynamic parameters of the molten pool include peak temperature, temperature gradient and cooling rate.

[0009] In an optional embodiment, According to the real-time detection of the surface morphology of the part by the visual sensor, the defect characteristic parameter vector is extracted and added to the pre-trained physical drive model, the defect influencing factors are inverted in combination with the digital twin model, the defect cause probability distribution is obtained, and the defect risk level is obtained in combination with the dynamic evolution trend prediction, the defect cause probability distribution, the defect characteristic parameter vector and the defect risk level are used as input, a dynamic control curve is generated and discretized into a multi-dimensional control instruction vector, the electromagnetic field parameters and the laser parameters are updated based on the multi-dimensional control instruction vector and the pre-set high-speed feedback channel, the initial control model parameters are generated in combination with the manufacturing task and the material type through the meta-learning algorithm, and the pre-set control model is updated according to the dynamic control curve, the optimized control model parameters are output and the crack suppression operation is performed again, including: Based on a pre-set visual sensor, real-time imaging of the produced parts is performed to obtain the surface morphology of the parts. The defect area in the surface morphology of the parts is extracted through an image processing algorithm and morphological characteristic parameters are calculated to obtain a defect characteristic parameter vector. The defect characteristic parameter vector is added to a pre-trained physical driving model. Based on the finite element method and material properties, a physical equation corresponding to the defect evolution is established, and the physical equation is solved to obtain the dynamic evolution trend of the defect. A digital twin model of the part manufacturing process is constructed based on the historical process data acquired in advance, and the causes of defects are inversely analyzed based on the digital twin model, and the inversion results are generated and merged with the dynamic evolution trend of the defects to obtain the probability distribution of the causes of defects, and the defect risk level is calculated based on the probability distribution of the causes of defects and the dynamic evolution trend of the defects, and the defect cause probability distribution, the defect characteristic parameter vector and the defect risk level are used as inputs to generate a dynamic control curve according to a deep reinforcement learning algorithm; The defect suppression process is modeled as a Markov decision process, with defect characteristic parameter vectors and risk levels as state spaces, and electromagnetic field parameters and laser parameters as action spaces. A defect suppression reward function is constructed according to the defect size reduction and the defect risk level reduction. The objective function corresponding to the Markov decision process is solved to obtain the optimal control strategy and represent the time-varying parameter curve. The state space is adaptively divided through an active learning strategy and the dynamic control curve is discretized into a multi-dimensional control instruction vector. The electromagnetic field parameters and laser parameters are updated through a pre-set high-speed feedback channel. Based on the meta-learning algorithm, combined with the goal of the manufacturing task and the corresponding feature vector of the material type, the initial control model parameters corresponding to the preset control model are generated through the few-sample learning method. The long-term reward loss function is minimized based on the dynamic control curve and the gradient descent algorithm to solve the optimized control model parameters. Based on the optimized control model parameters, new electromagnetic field parameters and laser parameters are generated. The control instructions of the manufacturing equipment are updated in real time through the preset process feedback channel. For the electromagnetic field parameters, the electromagnetic field distribution is dynamically regulated by adjusting the coil current amplitude and phase, the electromagnetic force distribution in the molten pool is changed, and crack propagation is suppressed. For the laser parameters, the energy input is dynamically controlled by adjusting the laser power, frequency and scanning path, the temperature gradient and cooling rate are optimized, the residual stress is reduced, and high-quality parts after crack suppression are obtained.

[0010] In an optional embodiment, The objective function corresponding to the Markov decision process is shown in the following formula: ; in, Indicates status s The optimal function value under s Indicates the current status of the system. a represents the action taken in state s, r ( s , a ) indicates that in the state s Take action a Rewards received, i represents the discount factor, Indicates in status s Take action a After transfer to state s' The probability of Indicates status s' The optimal function value under s' Indicates the system status after the state transition.

[0011] A second aspect of an embodiment of the present invention provides an electromagnetically assisted nickel-based alloy laser melting deposition crack suppression system, comprising: The first unit is used to construct an intelligent control system corresponding to electromagnetic field-assisted laser melting deposition, deeply fuse the pre-collected multi-source heterogeneous sensor data according to the multi-scale convolutional neural network in the intelligent control system, extract the high-dimensional feature vector corresponding to the dynamic evolution of the molten pool, and combine the long short-term memory network to make real-time predictions on the molten pool morphology, temperature distribution and stress distribution, generate prediction results and add the prediction results as the molten pool state vector to the central control unit, and the central control unit sets a reward function and an action-value network with crack tendency as a constraint according to the deep reinforcement learning algorithm, solves the optimized control strategy and parses it into execution instructions output to the laser and the pre-set electromagnetic coil array; The second unit is used to construct a three-dimensional electromagnetic field distribution matrix based on the multi-layer electromagnetic coils in the electromagnetic coil array, set orthogonal electromagnetic coils before and after the laser scanning direction and generate a first electromagnetic field vector, combine the real-time feedback of the molten pool morphology, dynamically generate an electromagnetic field parameter vector through a generative adversarial network and resolve it into a multi-dimensional control signal, send the multi-dimensional control signal to the electromagnetic coil array, combine with the induced current, stimulate the directional flow and nonlinear oscillation of the molten pool according to the resonance effect, obtain the oscillation parameters and predict the corresponding thermodynamic behavior of the molten pool according to the oscillation parameters, feed back the thermodynamic prediction results to the central control unit, and optimize and produce the laser parameters and electromagnetic field; The third unit is used to detect the surface morphology of parts in real time according to the visual sensor, extract the defect characteristic parameter vector and add it to the pre-trained physical driving model, invert the defect influencing factors in combination with the digital twin model, obtain the defect cause probability distribution and obtain the defect risk level in combination with the dynamic evolution trend prediction, take the defect cause probability distribution, the defect characteristic parameter vector and the defect risk level as input, generate a dynamic control curve and discretize it into a multi-dimensional control instruction vector, update the electromagnetic field parameters and laser parameters based on the multi-dimensional control instruction vector and the pre-set high-speed feedback channel, generate initial control model parameters in combination with the manufacturing task and material type through the meta-learning algorithm, update the pre-set control model according to the dynamic control curve, output the optimized control model parameters and perform crack suppression operation again.

[0012] According to a third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0013] According to a fourth aspect of the embodiments of the present invention, A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0014] In the present invention, multi-source sensor data are fused through a multi-scale convolutional neural network to extract high-dimensional feature vectors of the dynamic evolution of the molten pool, providing an information basis for subsequent state prediction and control decisions, and a long short-term memory network is used to predict the molten pool morphology, temperature distribution and stress distribution in real time. The prediction results are fed back to the central control unit as state vectors to achieve closed-loop control and dynamic optimization, improve the response speed and control accuracy of the system, and combine the molten pool morphology feedback. A generative adversarial network is used to dynamically generate electromagnetic field parameters to stimulate directional flow and nonlinear oscillation of the molten pool, improve the heat and mass transfer conditions of the molten pool, and improve the forming quality. The thermodynamic behavior of the molten pool is predicted according to the molten pool oscillation parameters, and the prediction results are fed back to the central control unit. The laser parameters and electromagnetic field parameters are optimized in real time to achieve precise control of temperature gradients, cooling rates, etc., and reduce residual stress and deformation. In summary, the present invention realizes all-round intelligent control of the laser melting deposition process, effectively improves manufacturing efficiency and part quality, reduces defect incidence and trial and error costs, has significant technical innovation and engineering application value, and provides new ideas and methods for intelligent upgrading in the field of additive manufacturing. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a schematic diagram of the process of the electromagnetic-assisted nickel-based alloy laser melting deposition crack suppression method according to an embodiment of the present invention; Figure 2 It is a schematic structural diagram of an electromagnetically assisted nickel-based alloy laser melting deposition crack suppression system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0016] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0017] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0018] Figure 1 FIG. 1 is a schematic flow chart of a method for suppressing cracks in a nickel-based alloy by electromagnetic-assisted laser melting deposition according to an embodiment of the present invention. Figure 1As shown, the method includes: S1. Construct an intelligent control system corresponding to electromagnetic field-assisted laser melting deposition, deeply fuse the pre-collected multi-source heterogeneous sensor data according to the multi-scale convolutional neural network in the intelligent control system, extract the high-dimensional feature vector corresponding to the dynamic evolution of the molten pool, combine the long short-term memory network to make real-time predictions on the molten pool morphology, temperature distribution and stress distribution, generate prediction results and add the prediction results as the molten pool state vector to the central control unit, the central control unit sets a reward function and an action-value network with crack tendency as a constraint according to the deep reinforcement learning algorithm, solves the optimized control strategy and parses it into execution instructions output to the laser and the pre-set electromagnetic coil array; The multi-scale convolutional neural network is a deep learning architecture that extracts features through convolution kernels of different scales to adapt to the various spatial scales and structures in the input data and enhance the perception ability of the model. The multi-source heterogeneous sensor data is diversified data from different types of sensors, which differ in format, dimension and sampling frequency, and is usually used for comprehensive analysis and decision support. The liquid metal area in the molten pool laser melting deposition process is key to the heat transfer, flow and final molding quality of the material. The molten pool morphology is used to describe the geometric characteristics of the molten pool, such as shape, depth and width, which has a direct impact on the quality of the final product. The molten pool state vector is a set of characteristic variables that describe the current state of the molten pool, which may include information such as temperature, speed, shape, etc., to help the model understand the dynamic changes of the molten pool. The crack tendency refers to the risk of cracking in the material under specific conditions. The action-value network is a model in reinforcement learning, which helps the intelligent agent select the best action by learning the value function of different actions in a specific state.

[0019] In an optional embodiment, An intelligent control system corresponding to electromagnetic field-assisted laser melting deposition is constructed. The multi-source heterogeneous sensor data collected in advance are deeply integrated according to the multi-scale convolutional neural network in the intelligent control system, and the high-dimensional feature vector corresponding to the dynamic evolution of the molten pool is extracted. The molten pool morphology, temperature distribution and stress distribution are predicted in real time in combination with the long short-term memory network, and the prediction results are generated and added to the central control unit as the molten pool state vector. The central control unit sets a reward function and an action-value network constrained by crack tendency according to the deep reinforcement learning algorithm, solves the optimized control strategy and parses it into execution instructions output to the laser and the pre-set electromagnetic coil array, including: The multi-source heterogeneous data corresponding to the electromagnetic field assisted laser melting deposition process is collected by pre-set sensors, wherein the multi-source heterogeneous data includes temperature field data, molten pool morphology image and material data, and an intelligent control system is constructed. For the temperature field data, the multi-receptive field convolution kernel corresponding to the multi-scale convolutional neural network in the intelligent control system is divided into multiple regions of interest. For each region of interest, the multi-receptive field convolution kernel is used to extract features to obtain a molten pool temperature distribution feature matrix. The molten pool morphology image is divided into grids of different scales by a spatial pyramid pooling algorithm, and the image features in the grid are extracted by a multi-scale convolution kernel to obtain a multi-scale feature vector of the molten pool morphology. The multi-scale feature vector of the molten pool morphology is deeply fused in combination with a cross-modal attention mechanism to obtain the high-dimensional feature vector. Based on the high-dimensional feature vector, the spatial features corresponding to the molten pool morphology are sequentially modeled through the convolutional long short-term memory network to obtain the spatiotemporal evolution law of the molten pool morphology, the molten pool temperature distribution is encoded and decoded through the encoding and decoding long short-term memory network to generate a temperature distribution map at a future moment, the stress field is weightedly modeled through the attention long and short time mechanism to obtain the stress distribution, and the spatiotemporal evolution law, temperature distribution map and stress distribution are combined to obtain the prediction result; The prediction result is added to the central control unit as a molten pool state vector. The central control unit constructs a molten pool thermal-fluid coupling model according to the heat conduction equation and the fluid dynamics equation, generates an initial control strategy and uses it as prior knowledge. The deep reinforcement learning algorithm is combined to perform data-driven strategy optimization. The crack tendency is used as a constraint condition, a reward function is constructed and an action-value network is generated. The reward function is solved to obtain an optimized control strategy, and the optimized control strategy is added to a pre-set laser and electromagnetic coil array.

[0020] The temperature field data is the temperature distribution information of different points in the molten pool area, reflecting the dynamic changes of heat conduction and melting process. The molten pool morphology image is a visual representation of the geometric features of the molten pool captured by imaging technology, which is used for quality analysis. The region of interest is a specific area of ​​concern in data analysis, which usually contains key features. The molten pool temperature distribution feature matrix is ​​structured data obtained after processing the molten pool temperature field, which is convenient for subsequent analysis. The spatial pyramid pooling algorithm is a technology that enhances the model's adaptability to scale changes by performing pooling operations at different scales. The cross-modal attention mechanism is used to integrate information between different modalities, thereby improving the model's ability to understand multi-source data. The sequence modeling is a processing The technology for processing and predicting time series data helps capture dynamic characteristics that change over time. The spatiotemporal evolution law describes the changing pattern of physical phenomena in time and space, which helps to understand the behavior of complex systems. The encoding and decoding long short-term memory network is a neural network structure for sequence-to-sequence tasks, including an encoder and a decoder, and is suitable for applications such as machine translation. The heat conduction equation is a partial differential equation that describes the heat transfer inside the material, indicating the change of temperature with time and space. The fluid dynamics equation describes the motion law of the fluid, involving the conservation of mass, momentum and energy. The molten pool heat-fluid coupling model comprehensively considers the heat conduction and fluid flow of the molten pool, and describes the interaction between heat and fluid in the laser melting process.

[0021] An intelligent control system corresponding to electromagnetic field-assisted laser melting deposition is constructed. The system includes a multi-source heterogeneous data acquisition module, a multi-scale convolutional neural network module, a long short-term memory network module and a central control unit. The data acquisition module collects multi-source heterogeneous data in the electromagnetic field-assisted laser melting deposition process through pre-set sensors, including temperature field data, molten pool morphology images and material data. The temperature field data is collected by an infrared thermal imager with a sampling frequency of 100 Hz and a spatial resolution of 0.1 mm. The molten pool morphology image is collected by a high-speed camera with a frame rate of 1000 fps and a resolution of 1024×1024 pixels. The material data, including powder particle size distribution, chemical composition, etc., are collected by a laser particle size analyzer and a spectrometer.

[0022] The multi-scale convolutional neural network module deeply fuses the collected multi-source heterogeneous data and extracts the high-dimensional feature vector corresponding to the dynamic evolution of the molten pool. For the temperature field data, three convolution kernels of 3×3, 5×5, and 7×7 are used to divide the temperature field into 9 regions of interest. Each region is extracted with three convolution kernels of different scales to obtain 27 feature maps, which are then reduced in dimension by 1×1 convolution to finally obtain a 9×64-dimensional molten pool temperature distribution feature matrix. For the molten pool morphology image, the spatial pyramid pooling algorithm is used to divide the image into three grids of 1×1, 2×2, and 4×4. Two convolution kernels of 3×3 and 5×5 are applied to each grid to extract features to obtain 21 feature maps. The 21×128-dimensional molten pool morphology multi-scale feature vector is obtained by global average pooling. Then, the temperature distribution feature matrix and the morphology multi-scale feature vector are weighted fused through the cross-modal attention mechanism to obtain a 256-dimensional high-dimensional feature vector.

[0023] The long short-term memory network module predicts the molten pool morphology, temperature distribution and stress distribution in real time based on high-dimensional feature vectors. For the molten pool morphology, a three-layer convolutional LSTM network is used. The input is a high-dimensional feature vector of 5 consecutive frames. The number of hidden layer neurons is 128, and the output is a sequence of molten pool contour coordinates for the next 3 frames. For temperature distribution, an encoding and decoding LSTM network is used. The encoder encodes the temperature distribution feature matrix of 5 consecutive frames into a 128-dimensional vector, and the decoder decodes the vector into a temperature distribution map for the next 3 frames. For stress distribution, an attention LSTM network is used. The input is a high-dimensional feature vector of 5 consecutive frames. The features of different time steps are weighted through a three-layer attention mechanism, and the output is a stress distribution map for the next 3 frames.

[0024] The central control unit receives the prediction results of the long short-term memory network module as the molten pool state vector, and generates an optimized control strategy based on the deep reinforcement learning algorithm. First, the molten pool thermal-fluid coupling model is constructed according to the heat conduction equation and the fluid dynamics equation to generate the initial control strategy. Then, a reward function with crack tendency as a constraint is set, and a 4-layer fully connected network is constructed as an action-value network. The input is the molten pool state vector, and the output is the value estimate of different control actions. The deep Q learning algorithm is used to optimize the parameters of the action-value network through multiple iterations, and finally the optimized control strategy is obtained.

[0025] The optimization control strategy includes parameters such as laser power, scanning speed, electromagnetic field strength and frequency. The central control unit interprets these parameters into specific execution instructions and sends them to the laser and the pre-set electromagnetic coil array. The laser adjusts the power and scanning path according to the instructions, and the electromagnetic coil array adjusts the current intensity and frequency according to the instructions, thereby achieving precise control of the molten pool state.

[0026] For example, taking TC4 titanium alloy as an example, when the crack tendency is detected to exceed the threshold, the optimized control strategy will reduce the laser power from 2000W to 1800W, increase the scanning speed from 10mm / s to 12mm / s, and increase the electromagnetic field intensity from 0.2T to 0.25T and the frequency from 50Hz to 60Hz.

[0027] In this embodiment, by collecting temperature field data, molten pool morphology images and material data, the system can comprehensively capture various information in the deposition process, and provide a more comprehensive and accurate description of the process state. The multi-scale feature extraction method can capture various details from micro to macro, and improve the richness and accuracy of feature representation. Using convolutional long short-term memory networks and encoding and decoding long short-term memory networks, the system can model the spatiotemporal evolution of the molten pool, which can not only understand the current state, but also predict the future state, providing an important basis for the formulation of control strategies. The combination of physical models and data-driven methods not only ensures the physical rationality of the control strategy, but also improves its adaptability and optimization. In summary, this embodiment is expected to significantly improve the quality, efficiency and reliability of the laser melting deposition process, which is of great significance for the manufacture of high-precision and high-performance components.

[0028] In an optional embodiment, Combined with the deep reinforcement learning algorithm for data-driven strategy optimization, the crack tendency is used as a constraint condition, and the reward function is constructed as shown in the following formula: ; in, R Represents the reward value, l represents the crack tendency proportionality coefficient, e represents the strain at the current molten pool, e c represents the critical strain, β represents the temperature gradient proportionality coefficient, N Indicates the number of points in the molten pool, ▽ T i Indicates i The temperature gradient at a point, c represents the molten pool shape proportional coefficient, l represents the actual length of the molten pool, l 0 represents the target length of the molten pool, w Indicates the actual width of the molten pool, w 0 represents the target width of the molten pool, s Represents the shape error tolerance.

[0029] In this embodiment, by incorporating crack tendency into the reward function, the formation of cracks can be effectively controlled and reduced, and optimizing the temperature gradient can reduce thermal stress and improve cooling uniformity, thereby improving the microstructure and mechanical properties. Comparing the actual molten pool size with the target size helps to ensure the uniformity and accuracy of the deposited layer and improve the geometric accuracy of parts. The reward function comprehensively considers three aspects: crack control, temperature distribution and molten pool shape, and realizes multi-objective optimization, which helps to maintain good processing efficiency while improving processing quality. In summary, this embodiment helps to significantly improve the stability, controllability and final product quality of the laser melting deposition process, and provides a powerful tool for achieving high-quality and efficient metal additive manufacturing.

[0030] S2. Construct a three-dimensional electromagnetic field distribution matrix based on the multi-layer electromagnetic coils in the electromagnetic coil array, set orthogonal electromagnetic coils before and after the laser scanning direction and generate a first electromagnetic field vector, combine the real-time feedback of the molten pool morphology, dynamically generate an electromagnetic field parameter vector through a generative adversarial network and resolve it into a multi-dimensional control signal, send the multi-dimensional control signal to the electromagnetic coil array, combine with the induced current, stimulate the directional flow and nonlinear oscillation of the molten pool according to the resonance effect, obtain the oscillation parameters and predict the corresponding thermodynamic behavior of the molten pool according to the oscillation parameters, feed back the thermodynamic prediction results to the central control unit, and optimize and produce the laser parameters and electromagnetic field; The multi-layer electromagnetic coil is a structure composed of multiple coils stacked together, and is usually used to enhance the intensity and distribution effect of the electromagnetic field. The three-dimensional electromagnetic field distribution matrix is ​​a mathematical model that describes the intensity and direction of the electromagnetic field at each point in three-dimensional space, which is convenient for analyzing the characteristics of the electromagnetic field. The first electromagnetic field vector represents the basic characteristics of the electromagnetic field, and usually includes the intensity and direction of the electric field and the magnetic field. The multi-dimensional control signal is a signal combination that jointly affects the behavior of the system through multiple input signals, and is suitable for the control of complex systems. The resonance effect refers to the phenomenon that the system exhibits a significantly enhanced response at a specific frequency, and is often used to optimize the performance of the electromagnetic system. The directional flow of the molten pool refers to the controlled flow direction of the fluid in the molten pool, thereby affecting the heat transfer and quality of the material. The oscillation parameters are a set of indicators that describe the oscillation characteristics of the system, such as frequency and amplitude, reflecting its dynamic behavior. The thermodynamic behavior describes the properties and change laws of the material during heat transfer and phase change.

[0031] In an optional embodiment, A three-dimensional electromagnetic field distribution matrix is ​​constructed based on the multi-layer electromagnetic coils in the electromagnetic coil array, orthogonal electromagnetic coils are arranged before and after the laser scanning direction to generate a first electromagnetic field vector, and an electromagnetic field parameter vector is dynamically generated by a generative adversarial network in combination with real-time feedback of the molten pool morphology and parsed into a multi-dimensional control signal, and the multi-dimensional control signal is sent to the electromagnetic coil array, including: Based on the multi-layer electromagnetic coils in the electromagnetic coil array, a three-dimensional electromagnetic field distribution matrix is ​​constructed by three-dimensional modeling software, orthogonal coils are arranged in the front and rear directions of the electromagnetic coil array, and the electromagnetic field vector is adjusted by adjusting the current phase and amplitude to obtain the first electromagnetic field vector; A molten pool morphology image is acquired by a pre-set high-speed camera, the acquired molten pool morphology image is pre-processed, the molten pool edge contour is extracted, the molten pool area is calculated by combining the semi-major axis and the semi-minor axis corresponding to the molten pool obtained by fitting the elliptic curve, the depth of the molten pool is estimated by combining the grayscale change, a neural network model is constructed and the current electromagnetic field parameters are used as input, and the molten pool characteristic parameters composed of the molten pool area, the molten pool depth and the molten pool edge contour are used as output, a mapping relationship is fitted and a feedback signal is generated; The generator in the generative adversarial network receives random noise and the feedback signal, dynamically generates an electromagnetic field parameter vector in combination with the first electromagnetic field vector, predicts the corresponding molten pool state and compares it with the feedback signal, determines the adversarial loss and optimizes the parameters of the generator, repeats generation and discrimination, uses the electromagnetic field parameter vector as the action space of reinforcement learning, uses the deviation between the molten pool characteristic parameters and the target parameters as the reward function, updates the weights of the generative adversarial network in combination with the gradient strategy and generates the optimal electromagnetic field parameters, and parses the optimal electromagnetic field parameters into multidimensional control signals and sends them to the electromagnetic coil array.

[0032] The orthogonal coil refers to an electromagnetic coil arranged perpendicular to each other in space, and is usually used to homogenize the magnetic field. The molten pool edge profile is the external shape feature of the molten pool surface, which affects the melting and cooling process of the material. The fitted elliptical curve is a mathematical method used to represent the molten pool edge data as an elliptical shape in order to analyze its geometric characteristics. The semi-major axis and semi-minor axis are the two main parameters describing the ellipse, which represent the longest and shortest diameters of the ellipse respectively. The molten pool characteristic parameters are indicators used to quantitatively describe the shape and state of the molten pool, such as depth, width and shape factor.

[0033] Based on the multi-layer electromagnetic coils in the electromagnetic coil array, a three-dimensional electromagnetic field distribution matrix is ​​constructed using three-dimensional modeling software such as COMSOL Multiphysics. Orthogonal coils are arranged in the front and rear directions of the electromagnetic coil array, and the electromagnetic field vector is adjusted by adjusting the current phase and amplitude. For example, coils can be arranged in the X, Y, and Z directions respectively, and the electromagnetic field strength and direction can be adjusted by changing the current size and phase of the coils in the three directions to obtain the first electromagnetic field vector. Specifically, the coil current in the X direction can be set to 1A and a phase of 0°, the coil current in the Y direction can be set to 0.8A and a phase of 90°, and the coil current in the Z direction can be set to 0.6A and a phase of 180°, thereby generating an electromagnetic field vector inclined at 45°.

[0034] The molten pool morphology image is obtained by a pre-set high-speed camera. The collected image is preprocessed, including grayscale, binarization, edge detection and other steps to extract the edge contour of the molten pool. Then, by fitting the elliptic curve, the semi-major axis and semi-minor axis corresponding to the molten pool are obtained, and the molten pool area is calculated. At the same time, the depth of the molten pool is estimated in combination with the grayscale change of the image. For example, a corresponding relationship can be established between the grayscale value and the depth, and the lower the grayscale value, the greater the depth. A neural network model containing multiple hidden layers is constructed, with the current electromagnetic field parameters (such as the current magnitude and phase in the three directions mentioned above) as input, and the molten pool characteristic parameters composed of the molten pool area, depth and edge contour as output, fitting the mapping relationship between them and generating a feedback signal.

[0035] Design a generative adversarial network, in which the generator receives random noise and the above feedback signal as input. The generator dynamically generates an electromagnetic field parameter vector, predicts the corresponding molten pool state and compares it with the feedback signal, calculates the adversarial loss and optimizes the parameters of the generator. Repeat the generation and discrimination process to continuously improve the performance of the generator. The generated electromagnetic field parameter vector is used as the action space of reinforcement learning, and the deviation between the molten pool characteristic parameters and the target parameters is used as the reward function. For example, the negative value of the mean square error between the molten pool area, depth and the target value can be used as a reward. Combined with the gradient strategy, the weights of the generative adversarial network are updated to generate the optimal electromagnetic field parameters.

[0036] The optimal electromagnetic field parameters are parsed into multi-dimensional control signals. For example, the electromagnetic field strength and direction can be decomposed into components in the three directions of X, Y, and Z, and converted into the current magnitude and phase signal of the corresponding coil. These multi-dimensional control signals are sent to the electromagnetic coil array to achieve precise control of the molten pool morphology. The whole process forms a closed-loop control system, which continuously adjusts the electromagnetic field parameters according to the real-time status of the molten pool to achieve the purpose of optimizing the molten pool morphology.

[0037] In this embodiment, a mapping relationship between the electric field parameters and the molten pool characteristics is established through a neural network model, which can capture complex nonlinear relationships and improve the accuracy of prediction. The use of generative adversarial networks and reinforcement learning can quickly respond to changes in the molten pool state and continuously optimize the control effect. Multi-objective optimization can comprehensively improve the quality and accuracy of the deposition process. It can not only respond to the current state, but also predict the future state and make corresponding adjustments, thereby realizing predictive control. In summary, this embodiment can significantly improve the performance and reliability of the laser melting deposition process, and is particularly suitable for the production of high-performance and high-precision metal parts, representing an important trend in the development of additive manufacturing technology towards intelligence and precision.

[0038] In an optional embodiment, Combined with the induced current, the directional flow and nonlinear oscillation of the molten pool are excited according to the resonance effect, the oscillation parameters are obtained and the thermodynamic behavior corresponding to the molten pool is predicted according to the oscillation parameters, the thermodynamic prediction results are fed back to the central control unit, and the laser parameters and electromagnetic fields are optimized and produced, including: Based on the multi-dimensional control signal and the induced current in the electromagnetic field, the volume electromagnetic force and the skin depth corresponding to the electromagnetic field frequency are calculated, the natural oscillation frequency of the molten pool is calculated based on the molten pool morphology, the flow and heat transfer behavior of the molten pool is described according to the fluid-solid coupling model, the boundary conditions are set and the control equations are discretized in combination with the preferential volume method, the physical quantities are stored in a staggered grid and the second-order upwind format is used for the flow direction, the diffusion term is second-order differentiated, the molten pool oscillation amplitude and oscillation frequency characteristics are solved, the molten pool oscillation amplitude and the oscillation frequency characteristics are combined into oscillation parameters and output; Based on the oscillation parameters, an electromagnetic induced oscillation experiment is carried out according to the induced current and the resonance effect, and a mapping model between the oscillation parameters and the thermodynamic parameters of the molten pool is constructed. The pre-set nonlinear regression model is trained in combination with the machine learning algorithm to obtain a prediction result, wherein the prediction result is the thermodynamic behavior corresponding to the molten pool, and the prediction result is fed back to the central control unit. By comparing the prediction result with the pre-set target result, an error value is generated, and the laser parameters and the electromagnetic field are optimized according to the error value, and production is carried out according to the optimized parameters, wherein the thermodynamic parameters of the molten pool include peak temperature, temperature gradient and cooling rate.

[0039] The volume electromagnetic force refers to the electromagnetic force acting within the volume of an object, affecting the movement and thermal state of the object. The skin depth describes the depth of the current distribution in the conductor, which is usually related to the frequency. The natural oscillation frequency is the frequency of the natural oscillation of the system in the absence of external force, reflecting its inherent characteristics. The fluid-solid coupling model is a mathematical model that describes the interaction between fluid and solid, and is suitable for analyzing complex flow and deformation behaviors. The preferential volume method is a numerical method that solves flow and heat transfer problems by controlling the priority of the volume. The second-order upwind scheme is a numerical discretization method used to improve the accuracy of calculations, and is particularly suitable for solving convection problems.

[0040] Based on the multi-dimensional control signal and the induced current in the electromagnetic field, the skin depth corresponding to the volume electromagnetic force and the electromagnetic field frequency is calculated. Specifically, by measuring the magnitude and frequency of the induced current in the electromagnetic field, substituting it into the Lorentz force formula, the volume electromagnetic force is calculated. For example, when the induced current is 100A and the frequency is 10kHz, the calculated volume electromagnetic force is 0.5N / m 3 At the same time, the corresponding skin depth is calculated according to the electromagnetic field frequency and the material conductivity. For example, for aluminum alloy material, the skin depth is about 0.8mm at a frequency of 10kHz.

[0041] The natural oscillation frequency of the molten pool is calculated based on the molten pool morphology. The molten pool morphology image is acquired in real time by a high-speed camera, and the geometric parameters such as the length, width and height of the molten pool are extracted. The natural oscillation frequency of the molten pool is calculated by substituting it into the Rayleigh formula. For example, when the length of the molten pool is 3mm, the width is 2mm, and the depth is 1mm, the calculated natural oscillation frequency is about 1kHz.

[0042] The flow and heat transfer behavior of the molten pool are described according to the fluid-solid coupling model. A multiphase flow model containing the liquid, solid and gas phases of the molten pool is established, and the VOF method is used to track the free surface, taking into account factors such as surface tension, buoyancy, and Marangoni convection. Boundary conditions such as laser heat source and convection heat dissipation are set, and the control equations are discretized in combination with the finite volume method. A staggered grid is used to store physical quantities, a second-order upwind format is used for the flow direction, and a second-order central difference is performed on the diffusion term. The velocity field, temperature field, etc. are obtained by solving the SIMPLE algorithm.

[0043] Analyze the oscillation characteristics of the molten pool and extract the oscillation amplitude and frequency. Use FFT transformation to analyze the change of the surface displacement of the molten pool over time and obtain the main oscillation frequency and amplitude. For example, the analysis results show that the main oscillation frequency of the molten pool is 800Hz and the amplitude is 0.1mm. Use these oscillation parameters as output results.

[0044] Based on the above oscillation parameters, electromagnetic induced oscillation experiments were conducted in combination with induced current and resonance effect. By changing the induced current frequency, frequency sweep experiments were conducted in the range of 50-2000Hz with an interval of 100Hz, and the oscillation behavior of the molten pool at different frequencies was recorded. At the same time, the temperature field distribution of the molten pool was measured to obtain thermodynamic parameters such as peak temperature, temperature gradient and cooling rate.

[0045] Construct a mapping model between oscillation parameters and thermodynamic parameters of the molten pool. Take the oscillation frequency and amplitude as input, and the peak temperature, temperature gradient, and cooling rate as output, and use a nonlinear regression method to establish the mapping relationship. For example, a multilayer perceptron neural network can be used, with a hidden layer structure of 3 layers, each layer containing 10 neurons, and the activation function is the ReLU function.

[0046] Combine machine learning algorithms to train the pre-set nonlinear regression model. Use the experimentally obtained data set, 80% for training and 20% for validation. Use the Adam optimizer, set the learning rate to 0.001, and train for 500 epochs. Use the back-propagation algorithm to continuously adjust the network parameters and finally obtain a trained prediction model.

[0047] By using the trained model and inputting the current oscillation parameters, the corresponding thermodynamic behavior of the molten pool can be predicted, including the peak temperature, temperature gradient, and cooling rate. For example, when the input oscillation frequency is 800 Hz and the amplitude is 0.1 mm, the predicted peak temperature is 1800 K, the temperature gradient is 106 K / m, and the cooling rate is 10 4 K / s.

[0048] The prediction results are fed back to the central control unit and compared with the preset target results. Assuming the target peak temperature is 1750K and the temperature gradient is 1.2×10 6 K / m, cooling rate is 1.5×10 4 K / s. The error between the predicted result and the target value was calculated, and the temperature error was 50K and the temperature gradient error was 2×10 5 K / m, the cooling rate error is 5×10 3 K / s.

[0049] According to the calculated error value, the laser parameters and electromagnetic field are optimized. For example, the peak temperature can be increased by increasing the laser power by 5%, the temperature gradient can be increased by reducing the scanning speed by 10%, and the cooling rate can be increased by adjusting the electromagnetic field frequency to 850Hz. The specific optimization strategy can use the PID control algorithm to dynamically adjust each parameter according to the error size and change trend.

[0050] Production is carried out according to the optimized laser parameters and electromagnetic field parameters. The laser power is adjusted to 1050W, the scanning speed is 900mm / s, the electromagnetic field frequency is 850Hz, etc. During the production process, the oscillation behavior of the molten pool and the thermodynamic parameters are continuously monitored and continuously optimized to achieve closed-loop control.

[0051] In this embodiment, by calculating the volume electromagnetic force, skin depth and natural oscillation frequency of the molten pool, the system can accurately describe the dynamic behavior of the molten pool, providing a solid theoretical basis for subsequent control and optimization. The fluid-solid coupling model is used to describe the flow and heat transfer behavior of the molten pool, realizing a comprehensive analysis of the complex physical processes of the molten pool, and being able to more accurately predict the behavior of the molten pool. Through the preferential volume method, staggered grids and high-order difference formats, high-precision numerical simulation is achieved, ensuring the accuracy and reliability of the calculation results. By solving the oscillation amplitude and frequency characteristics of the molten pool, the dynamic behavior of the molten pool can be deeply understood. In summary, this embodiment can not only improve the processing quality and efficiency, but also realize the precise regulation of material properties, and provide strong technical support for innovative applications in aerospace, medical devices, high-end equipment manufacturing and other fields.

[0052] S3. Detect the surface morphology of parts in real time according to the visual sensor, extract the defect characteristic parameter vector and add it to the pre-trained physical driving model, and invert the defect influencing factors with the digital twin model to obtain the defect cause probability distribution and obtain the defect risk level in combination with the dynamic evolution trend prediction. The defect cause probability distribution, the defect characteristic parameter vector and the defect risk level are used as input to generate a dynamic control curve and discretize it into a multi-dimensional control instruction vector. The electromagnetic field parameters and laser parameters are updated based on the multi-dimensional control instruction vector and the pre-set high-speed feedback channel. Through the meta-learning algorithm, the initial control model parameters are generated in combination with the manufacturing task and the material type, and the pre-set control model is updated according to the dynamic control curve. The optimized control model parameters are output and the crack suppression operation is performed again.

[0053] The dynamic control curve is a curve that describes how the system adjusts and optimizes operations during time changes, and is used for real-time control and optimization decisions. The multidimensional control instruction vector is a vector containing multiple control instructions, which is used to simultaneously adjust multiple system parameters to achieve the expected goals. The meta-learning algorithm is a learning algorithm that aims to learn how to quickly adapt to new tasks through experience, and is often used to improve the performance of the model in small sample learning. The high-speed feedback channel refers to a channel for realizing rapid information feedback in the system, which is used to monitor and adjust the system status in real time.

[0054] In an optional embodiment, According to the real-time detection of the surface morphology of the part by the visual sensor, the defect characteristic parameter vector is extracted and added to the pre-trained physical drive model, the defect influencing factors are inverted in combination with the digital twin model, the defect cause probability distribution is obtained, and the defect risk level is obtained in combination with the dynamic evolution trend prediction, the defect cause probability distribution, the defect characteristic parameter vector and the defect risk level are used as input, a dynamic control curve is generated and discretized into a multi-dimensional control instruction vector, the electromagnetic field parameters and the laser parameters are updated based on the multi-dimensional control instruction vector and the pre-set high-speed feedback channel, the initial control model parameters are generated in combination with the manufacturing task and the material type through the meta-learning algorithm, and the pre-set control model is updated according to the dynamic control curve, the optimized control model parameters are output and the crack suppression operation is performed again, including: Based on a pre-set visual sensor, real-time imaging of the produced parts is performed to obtain the surface morphology of the parts. The defect area in the surface morphology of the parts is extracted through an image processing algorithm and morphological characteristic parameters are calculated to obtain a defect characteristic parameter vector. The defect characteristic parameter vector is added to a pre-trained physical driving model. Based on the finite element method and material properties, a physical equation corresponding to the defect evolution is established, and the physical equation is solved to obtain the dynamic evolution trend of the defect. A digital twin model of the part manufacturing process is constructed based on the historical process data acquired in advance, and the causes of defects are inversely analyzed based on the digital twin model, and the inversion results are generated and merged with the dynamic evolution trend of the defects to obtain the probability distribution of the causes of defects, and the defect risk level is calculated based on the probability distribution of the causes of defects and the dynamic evolution trend of the defects, and the defect cause probability distribution, the defect characteristic parameter vector and the defect risk level are used as inputs to generate a dynamic control curve according to a deep reinforcement learning algorithm; The defect suppression process is modeled as a Markov decision process, with defect characteristic parameter vectors and risk levels as state spaces, and electromagnetic field parameters and laser parameters as action spaces. A defect suppression reward function is constructed according to the defect size reduction and the defect risk level reduction. The objective function corresponding to the Markov decision process is solved to obtain the optimal control strategy and represent the time-varying parameter curve. The state space is adaptively divided through an active learning strategy and the dynamic control curve is discretized into a multi-dimensional control instruction vector. The electromagnetic field parameters and laser parameters are updated through a pre-set high-speed feedback channel. Based on the meta-learning algorithm, combined with the goal of the manufacturing task and the corresponding feature vector of the material type, the initial control model parameters corresponding to the preset control model are generated through the few-sample learning method. The long-term reward loss function is minimized based on the dynamic control curve and the gradient descent algorithm to solve the optimized control model parameters. Based on the optimized control model parameters, new electromagnetic field parameters and laser parameters are generated. The control instructions of the manufacturing equipment are updated in real time through the preset process feedback channel. For the electromagnetic field parameters, the electromagnetic field distribution is dynamically regulated by adjusting the coil current amplitude and phase, the electromagnetic force distribution in the molten pool is changed, and crack propagation is suppressed. For the laser parameters, the energy input is dynamically controlled by adjusting the laser power, frequency and scanning path, the temperature gradient and cooling rate are optimized, the residual stress is reduced, and high-quality parts after crack suppression are obtained.

[0055] The morphological feature parameters are parameters that describe the shape and structure of an object and are used to analyze and understand the geometric features of the object. The inversion analysis is an analysis method that infers the internal state or parameters of a system through observation results and is often used for parameter estimation and model correction. The long-term reward loss function is an objective function in reinforcement learning, which is used to measure the cumulative rewards obtained by the intelligent agent under a series of actions, thereby optimizing the strategy.

[0056] Based on the pre-set visual sensor, the produced parts are imaged in real time to obtain the surface morphology information of the parts. Specifically, a high-resolution industrial camera or 3D scanner and other equipment can be used to perform an all-round scan of the surface of the part to obtain high-precision three-dimensional point cloud data or two-dimensional image data of the surface of the part. For example, an industrial camera with a resolution of 4096x3000 pixels is used to continuously shoot the surface of the part at a frequency of 30 frames per second to obtain a high-definition image sequence of the surface of the part.

[0057] The defect area in the surface morphology of the part is extracted through image processing algorithms and the morphological characteristic parameters are calculated. Specifically, image segmentation methods such as edge detection and region growing can be used to identify the defect area, and then the defect characteristics can be quantified through morphological indicators such as area, perimeter, and roundness. For example, for an irregular crack with an area of ​​2mm^2 and a perimeter of 6mm, its characteristic parameters such as roundness of 0.7 and aspect ratio of 2.5 can be calculated. These characteristic parameters are combined into an n-dimensional vector to obtain the defect characteristic parameter vector.

[0058] The above defect characteristic parameter vector is added to the pre-trained physical drive model. This physical drive model is built based on material science principles and finite element methods, and can simulate the evolution of defects under different process parameters. The model input includes material properties, process parameters, and defect characteristics, and the output is the dynamic changes of defect size, stress distribution, etc. By solving the physical equations, the dynamic evolution trend of defects over time can be obtained. For example, it is predicted that the crack length will increase by 0.5mm after 100 hours, and the stress at the crack tip will increase by 20MPa.

[0059] A digital twin model of the parts manufacturing process is constructed based on the historical process data acquired in advance. The model integrates multi-domain information such as CAD / CAE / CAM, and can realize virtual simulation of the manufacturing process. The model is used to perform inverse analysis on the causes of defects, and combined with methods such as Bayesian inference, the probability distribution of various possible causes is generated. For example, the analysis shows that the causes of defects are: thermal stress concentration (60%), inclusions (30%), and pores (10%). The probability distribution is fused with the above-mentioned dynamic evolution trend of defects to obtain a comprehensive probability distribution of the causes of defects.

[0060] Based on the probability distribution of defect causes and dynamic evolution trends, the defect risk level is calculated using the fuzzy comprehensive evaluation method. Factors such as defect size, expansion rate, and stress level are used as evaluation indicators, and corresponding weight coefficients are set to ultimately obtain a quantitative score for defect risk. For example, the risk level can be divided into three levels: low (0-0.3), medium (0.3-0.7), and high (0.7-1.0).

[0061] The defect cause probability distribution, defect characteristic parameter vector and defect risk level are used as input, and a deep reinforcement learning algorithm is used to generate a dynamic control curve. Specifically, the defect suppression process is modeled as a Markov decision process, with defect characteristics and risk level as the state space, and electromagnetic field parameters and laser parameters as the action space. A defect suppression reward function is constructed, in which the reduction in defect size and the reduction in risk level are positive incentives, and energy consumption is negative penalty. The optimal control strategy is solved by algorithms such as deep Q networks, and a parameter adjustment curve that changes over time is obtained.

[0062] An active learning strategy is used to adaptively divide the state space and discretize the continuous dynamic control curve into a multi-dimensional control instruction vector. For example, the 100-hour control process is discretized into 1000 time steps, and each time step corresponds to a set of electromagnetic field and laser parameters. Specifically, it may include coil current (0-100A), magnetic field frequency (0-1000Hz), laser power (0-2000W), scanning speed (0-10m / s) and other dimensions. Through the pre-set high-speed feedback channel, these discretized control instructions are transmitted to the manufacturing equipment in real time to update the electromagnetic field parameters and laser parameters.

[0063] Based on the meta-learning algorithm, combined with the specific manufacturing task objectives and material types, the initial parameters of the control model are generated through a few-sample learning method. For example, for the welding task of high-strength steel, the initial model can be quickly built based on 5-10 sets of historical data. Then, based on the dynamic control curve generated above, the gradient descent algorithm is used to minimize the long-term reward loss function and iteratively optimize the control model parameters. Finally, an optimized control model that can adapt to the current task is obtained.

[0064] Based on the optimized control model parameters, new electromagnetic field parameters and laser parameter sequences are generated. These parameters are updated to the manufacturing equipment in real time through the preset process feedback channel. For electromagnetic field parameters, the electromagnetic field distribution can be dynamically regulated by adjusting the coil current amplitude (such as 50-80A) and phase (such as 0-360°), changing the electromagnetic force distribution in the molten pool, and thus inhibiting crack propagation. For laser parameters, the energy input can be dynamically controlled by adjusting the laser power (such as 500-1500W), frequency (such as 50-200Hz) and scanning path, optimizing the temperature gradient and cooling rate, and reducing residual stress.

[0065] In this embodiment, through visual sensors and image processing algorithms, the surface morphology of parts can be imaged and defects extracted in real time, and defect characteristic parameters can be quantified to provide a basis for subsequent defect evolution prediction, risk assessment and control decision-making. Deep reinforcement learning is used to model defect suppression as a Markov decision process. The optimal control strategy is obtained by solving the objective function and expressed as a time-varying parameter curve. The electromagnetic field and laser parameters can be adjusted in real time according to the dynamic changes of defect parameters and risk levels to achieve precise control of defects. Initial parameters are generated through few-sample learning, and in the process of defect suppression, the parameters are optimized online based on the dynamic control curve to improve the generalization ability and adaptability of the control model. In summary, this embodiment can detect defect parameters in real time, accurately predict defect evolution, generate the optimal control strategy, and adaptively optimize the control model, ultimately achieving precise suppression of defects and improvement of part quality, providing new ideas for additive manufacturing of high-quality components, with broad engineering application prospects, and of great significance for promoting the intelligent development of the manufacturing industry.

[0066] In an optional embodiment, The objective function corresponding to the Markov decision process is shown in the following formula: ; in, Indicates status s The optimal function value under s Indicates the current status of the system. a represents the action taken in state s, r ( s , a ) indicates that in the state s Take action a Rewards received, i represents the discount factor, Indicates in status s Take action a After transfer to state s' The probability of Indicates status s' The optimal function value under s' Indicates the system status after the state transition.

[0067] In this embodiment, by maximizing the long-term cumulative return, the globally optimal control strategy can be obtained, avoiding short-sighted behavior and local optimality. By introducing the state transition probability, the control strategy can adapt to the random changes of the system state and improve the stability and robustness of the strategy. Through value function approximation, the original problem can be converted into an iterative update of the value function, avoiding exhaustive strategy search, and improving the convergence speed and computational efficiency of the algorithm. In summary, this embodiment provides a theoretical basis for the mathematical modeling of defect control strategies, provides new ideas and methods for the intelligent optimization of complex manufacturing processes, and has broad application prospects.

[0068] Figure 2 FIG. 1 is a schematic diagram of the structure of the electromagnetic-assisted nickel-based alloy laser melting deposition crack suppression system according to an embodiment of the present invention. Figure 2 As shown, the system comprises: The first unit is used to construct an intelligent control system corresponding to electromagnetic field-assisted laser melting deposition, deeply fuse the pre-collected multi-source heterogeneous sensor data according to the multi-scale convolutional neural network in the intelligent control system, extract the high-dimensional feature vector corresponding to the dynamic evolution of the molten pool, and combine the long short-term memory network to make real-time predictions on the molten pool morphology, temperature distribution and stress distribution, generate prediction results and add the prediction results as the molten pool state vector to the central control unit, and the central control unit sets a reward function and an action-value network with crack tendency as a constraint according to the deep reinforcement learning algorithm, solves the optimized control strategy and parses it into execution instructions output to the laser and the pre-set electromagnetic coil array; The second unit is used to construct a three-dimensional electromagnetic field distribution matrix based on the multi-layer electromagnetic coils in the electromagnetic coil array, set orthogonal electromagnetic coils before and after the laser scanning direction and generate a first electromagnetic field vector, combine the real-time feedback of the molten pool morphology, dynamically generate an electromagnetic field parameter vector through a generative adversarial network and resolve it into a multi-dimensional control signal, send the multi-dimensional control signal to the electromagnetic coil array, combine with the induced current, stimulate the directional flow and nonlinear oscillation of the molten pool according to the resonance effect, obtain the oscillation parameters and predict the corresponding thermodynamic behavior of the molten pool according to the oscillation parameters, feed back the thermodynamic prediction results to the central control unit, and optimize and produce the laser parameters and electromagnetic field; The third unit is used to detect the surface morphology of parts in real time according to the visual sensor, extract the defect characteristic parameter vector and add it to the pre-trained physical driving model, invert the defect influencing factors in combination with the digital twin model, obtain the defect cause probability distribution and obtain the defect risk level in combination with the dynamic evolution trend prediction, take the defect cause probability distribution, the defect characteristic parameter vector and the defect risk level as input, generate a dynamic control curve and discretize it into a multi-dimensional control instruction vector, update the electromagnetic field parameters and laser parameters based on the multi-dimensional control instruction vector and the pre-set high-speed feedback channel, generate initial control model parameters in combination with the manufacturing task and material type through the meta-learning algorithm, update the pre-set control model according to the dynamic control curve, output the optimized control model parameters and perform crack suppression operation again.

[0069] According to a third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0070] According to a fourth aspect of the embodiments of the present invention, A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0071] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0072] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. Electromagnetic assisted nickel-based alloy laser melting deposition crack suppression method, characterized in that: include: Construct an intelligent control system corresponding to electromagnetic field-assisted laser melting deposition, deeply fuse the pre-collected multi-source heterogeneous sensor data according to the multi-scale convolutional neural network in the intelligent control system, extract the high-dimensional feature vector corresponding to the dynamic evolution of the molten pool, combine the long short-term memory network to make real-time predictions on the molten pool morphology, temperature distribution and stress distribution, generate prediction results and add the prediction results as the molten pool state vector to the central control unit, the central control unit sets a reward function and an action-value network with crack tendency as a constraint according to the deep reinforcement learning algorithm, solves the optimized control strategy and parses it into execution instructions output to the laser and the pre-set electromagnetic coil array; A three-dimensional electromagnetic field distribution matrix is ​​constructed based on the multi-layer electromagnetic coils in the electromagnetic coil array, orthogonal electromagnetic coils are set before and after the laser scanning direction to generate a first electromagnetic field vector, and in combination with real-time feedback of the molten pool morphology, an electromagnetic field parameter vector is dynamically generated through a generative adversarial network and parsed into a multi-dimensional control signal, and the multi-dimensional control signal is sent to the electromagnetic coil array, and in combination with the induced current, the directional flow and nonlinear oscillation of the molten pool are stimulated according to the resonance effect, the oscillation parameters are obtained, and the corresponding thermodynamic behavior of the molten pool is predicted according to the oscillation parameters, and the thermodynamic prediction results are fed back to the central control unit to optimize and produce the laser parameters and the electromagnetic field; The surface morphology of parts is detected in real time by visual sensors, and the defect characteristic parameter vector is extracted and added to the pre-trained physical drive model. The defect influencing factors are inverted in combination with the digital twin model to obtain the defect cause probability distribution and the defect risk level is obtained in combination with the dynamic evolution trend prediction. The defect cause probability distribution, the defect characteristic parameter vector and the defect risk level are used as input to generate a dynamic control curve and discretize it into a multidimensional control instruction vector. The electromagnetic field parameters and laser parameters are updated based on the multidimensional control instruction vector and a pre-set high-speed feedback channel. Through the meta-learning algorithm, the initial control model parameters are generated in combination with the manufacturing task and the material type, and the pre-set control model is updated according to the dynamic control curve. The optimized control model parameters are output and the crack suppression operation is performed again.

2. The method according to claim 1, characterized in that An intelligent control system corresponding to electromagnetic field-assisted laser melting deposition is constructed. The multi-source heterogeneous sensor data collected in advance are deeply integrated according to the multi-scale convolutional neural network in the intelligent control system, and the high-dimensional feature vector corresponding to the dynamic evolution of the molten pool is extracted. The molten pool morphology, temperature distribution and stress distribution are predicted in real time in combination with the long short-term memory network, and the prediction results are generated and added to the central control unit as the molten pool state vector. The central control unit sets a reward function and an action-value network constrained by crack tendency according to the deep reinforcement learning algorithm, solves the optimized control strategy and parses it into execution instructions output to the laser and the pre-set electromagnetic coil array, including: The multi-source heterogeneous data corresponding to the electromagnetic field assisted laser melting deposition process is collected by pre-set sensors, wherein the multi-source heterogeneous data includes temperature field data, molten pool morphology image and material data, and an intelligent control system is constructed. For the temperature field data, the multi-receptive field convolution kernel corresponding to the multi-scale convolutional neural network in the intelligent control system is divided into multiple regions of interest. For each region of interest, the multi-receptive field convolution kernel is used to extract features to obtain a molten pool temperature distribution feature matrix. The molten pool morphology image is divided into grids of different scales by a spatial pyramid pooling algorithm, and the image features in the grid are extracted by a multi-scale convolution kernel to obtain a multi-scale feature vector of the molten pool morphology. The multi-scale feature vector of the molten pool morphology is deeply fused in combination with a cross-modal attention mechanism to obtain the high-dimensional feature vector. Based on the high-dimensional feature vector, the spatial features corresponding to the molten pool morphology are sequentially modeled through the convolutional long short-term memory network to obtain the spatiotemporal evolution law of the molten pool morphology, the molten pool temperature distribution is encoded and decoded through the encoding and decoding long short-term memory network to generate a temperature distribution map at a future moment, the stress field is weightedly modeled through the attention long and short time mechanism to obtain the stress distribution, and the spatiotemporal evolution law, temperature distribution map and stress distribution are combined to obtain the prediction result; The prediction result is added to the central control unit as a molten pool state vector. The central control unit constructs a molten pool thermal-fluid coupling model according to the heat conduction equation and the fluid dynamics equation, generates an initial control strategy and uses it as prior knowledge. The deep reinforcement learning algorithm is combined to perform data-driven strategy optimization. The crack tendency is used as a constraint condition, a reward function is constructed and an action-value network is generated. The reward function is solved to obtain an optimized control strategy, and the optimized control strategy is added to a pre-set laser and electromagnetic coil array.

3. The method according to claim 2, characterized in that Combined with the deep reinforcement learning algorithm for data-driven strategy optimization, the crack tendency is used as a constraint condition, and the reward function is constructed as shown in the following formula: ； in, R Represents the reward value, λ represents the crack tendency proportionality coefficient, ε represents the strain at the current molten pool, ε c represents the critical strain, β represents the temperature gradient proportionality coefficient, N Indicates the number of points in the molten pool, ▽ T i Indicates i The temperature gradient at a point, γ represents the molten pool shape proportional coefficient, l represents the actual length of the molten pool, l 0 represents the target length of the molten pool, w Indicates the actual width of the molten pool, w 0 represents the target width of the molten pool, σ Represents the shape error tolerance.

4. The method according to claim 1, characterized in that A three-dimensional electromagnetic field distribution matrix is ​​constructed based on the multi-layer electromagnetic coils in the electromagnetic coil array, orthogonal electromagnetic coils are arranged before and after the laser scanning direction to generate a first electromagnetic field vector, and an electromagnetic field parameter vector is dynamically generated by a generative adversarial network in combination with real-time feedback of the molten pool morphology and parsed into a multi-dimensional control signal, and the multi-dimensional control signal is sent to the electromagnetic coil array, including: Based on the multi-layer electromagnetic coils in the electromagnetic coil array, a three-dimensional electromagnetic field distribution matrix is ​​constructed by three-dimensional modeling software, orthogonal coils are arranged in the front and rear directions of the electromagnetic coil array, and the electromagnetic field vector is adjusted by adjusting the current phase and amplitude to obtain the first electromagnetic field vector; A molten pool morphology image is acquired by a pre-set high-speed camera, the acquired molten pool morphology image is pre-processed, the molten pool edge contour is extracted, the molten pool area is calculated by combining the semi-major axis and the semi-minor axis corresponding to the molten pool obtained by fitting the elliptic curve, the depth of the molten pool is estimated by combining the grayscale change, a neural network model is constructed and the current electromagnetic field parameters are used as input, and the molten pool characteristic parameters composed of the molten pool area, the molten pool depth and the molten pool edge contour are used as output, a mapping relationship is fitted and a feedback signal is generated; The generator in the generative adversarial network receives random noise and the feedback signal, dynamically generates an electromagnetic field parameter vector in combination with the first electromagnetic field vector, predicts the corresponding molten pool state and compares it with the feedback signal, determines the adversarial loss and optimizes the parameters of the generator, repeats generation and discrimination, uses the electromagnetic field parameter vector as the action space of reinforcement learning, uses the deviation between the molten pool characteristic parameters and the target parameters as the reward function, updates the weights of the generative adversarial network in combination with the gradient strategy and generates the optimal electromagnetic field parameters, and parses the optimal electromagnetic field parameters into multidimensional control signals and sends them to the electromagnetic coil array.

5. The method according to claim 1, characterized in that Combined with the induced current, the directional flow and nonlinear oscillation of the molten pool are excited according to the resonance effect, the oscillation parameters are obtained and the thermodynamic behavior corresponding to the molten pool is predicted according to the oscillation parameters, the thermodynamic prediction results are fed back to the central control unit, and the laser parameters and electromagnetic fields are optimized and produced, including: Based on the multi-dimensional control signal and the induced current in the electromagnetic field, the volume electromagnetic force and the skin depth corresponding to the electromagnetic field frequency are calculated, the natural oscillation frequency of the molten pool is calculated based on the molten pool morphology, the flow and heat transfer behavior of the molten pool is described according to the fluid-solid coupling model, the boundary conditions are set and the control equations are discretized in combination with the preferential volume method, the physical quantities are stored in a staggered grid and the second-order upwind format is used for the flow direction, the diffusion term is second-order differentiated, the molten pool oscillation amplitude and oscillation frequency characteristics are solved, the molten pool oscillation amplitude and the oscillation frequency characteristics are combined into oscillation parameters and output; Based on the oscillation parameters, an electromagnetic induced oscillation experiment is carried out according to the induced current and the resonance effect, and a mapping model between the oscillation parameters and the thermodynamic parameters of the molten pool is constructed. The pre-set nonlinear regression model is trained in combination with the machine learning algorithm to obtain a prediction result, wherein the prediction result is the thermodynamic behavior corresponding to the molten pool, and the prediction result is fed back to the central control unit. By comparing the prediction result with the pre-set target result, an error value is generated, and the laser parameters and the electromagnetic field are optimized according to the error value, and production is carried out according to the optimized parameters, wherein the thermodynamic parameters of the molten pool include peak temperature, temperature gradient and cooling rate.

6. The method according to claim 1, characterized in that According to the real-time detection of the surface morphology of the part by the visual sensor, the defect characteristic parameter vector is extracted and added to the pre-trained physical drive model, the defect influencing factors are inverted in combination with the digital twin model, the defect cause probability distribution is obtained, and the defect risk level is obtained in combination with the dynamic evolution trend prediction, the defect cause probability distribution, the defect characteristic parameter vector and the defect risk level are used as input, a dynamic control curve is generated and discretized into a multi-dimensional control instruction vector, the electromagnetic field parameters and the laser parameters are updated based on the multi-dimensional control instruction vector and the pre-set high-speed feedback channel, the initial control model parameters are generated in combination with the manufacturing task and the material type through the meta-learning algorithm, and the pre-set control model is updated according to the dynamic control curve, the optimized control model parameters are output and the crack suppression operation is performed again, including: Based on a pre-set visual sensor, real-time imaging of the produced parts is performed to obtain the surface morphology of the parts. The defect area in the surface morphology of the parts is extracted through an image processing algorithm and morphological characteristic parameters are calculated to obtain a defect characteristic parameter vector. The defect characteristic parameter vector is added to a pre-trained physical driving model. Based on the finite element method and material properties, a physical equation corresponding to the defect evolution is established, and the physical equation is solved to obtain the dynamic evolution trend of the defect. A digital twin model of the part manufacturing process is constructed based on the historical process data acquired in advance, and the causes of defects are inversely analyzed based on the digital twin model, and the inversion results are generated and merged with the dynamic evolution trend of the defects to obtain the probability distribution of the causes of defects, and the defect risk level is calculated based on the probability distribution of the causes of defects and the dynamic evolution trend of the defects, and the defect cause probability distribution, the defect characteristic parameter vector and the defect risk level are used as inputs to generate a dynamic control curve according to a deep reinforcement learning algorithm; The defect suppression process is modeled as a Markov decision process, with defect characteristic parameter vectors and risk levels as state spaces, and electromagnetic field parameters and laser parameters as action spaces. A defect suppression reward function is constructed according to the defect size reduction and the defect risk level reduction. The objective function corresponding to the Markov decision process is solved to obtain the optimal control strategy and represent the time-varying parameter curve. The state space is adaptively divided through an active learning strategy and the dynamic control curve is discretized into a multi-dimensional control instruction vector. The electromagnetic field parameters and laser parameters are updated through a pre-set high-speed feedback channel. Based on the meta-learning algorithm, combined with the goal of the manufacturing task and the corresponding feature vector of the material type, the initial control model parameters corresponding to the preset control model are generated through the few-sample learning method. The long-term reward loss function is minimized based on the dynamic control curve and the gradient descent algorithm to solve the optimized control model parameters. Based on the optimized control model parameters, new electromagnetic field parameters and laser parameters are generated. The control instructions of the manufacturing equipment are updated in real time through the preset process feedback channel. For the electromagnetic field parameters, the electromagnetic field distribution is dynamically regulated by adjusting the coil current amplitude and phase, the electromagnetic force distribution in the molten pool is changed, and crack propagation is suppressed. For the laser parameters, the energy input is dynamically controlled by adjusting the laser power, frequency and scanning path, the temperature gradient and cooling rate are optimized, the residual stress is reduced, and high-quality parts after crack suppression are obtained.

7. The method according to claim 1, characterized in that The objective function corresponding to the Markov decision process is shown in the following formula: ； in, Indicates status s The optimal function value under s Indicates the current status of the system. a represents the action taken in state s, r ( s , a ) indicates that in the state s Take action a Rewards received, θ represents the discount factor, Indicates in status s Take action a After transfer to state s' The probability of Indicates status s' The optimal function value under s' Indicates the system status after the state transition.

8. An electromagnetically assisted nickel-based alloy laser melting deposition crack suppression system, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to construct an intelligent control system corresponding to electromagnetic field-assisted laser melting deposition, deeply fuse the pre-collected multi-source heterogeneous sensor data according to the multi-scale convolutional neural network in the intelligent control system, extract the high-dimensional feature vector corresponding to the dynamic evolution of the molten pool, and combine the long short-term memory network to make real-time predictions on the molten pool morphology, temperature distribution and stress distribution, generate prediction results and add the prediction results as the molten pool state vector to the central control unit, and the central control unit sets a reward function and an action-value network with crack tendency as a constraint according to the deep reinforcement learning algorithm, solves the optimized control strategy and parses it into execution instructions output to the laser and the pre-set electromagnetic coil array; The second unit is used to construct a three-dimensional electromagnetic field distribution matrix based on the multi-layer electromagnetic coils in the electromagnetic coil array, set orthogonal electromagnetic coils before and after the laser scanning direction and generate a first electromagnetic field vector, combine the real-time feedback of the molten pool morphology, dynamically generate an electromagnetic field parameter vector through a generative adversarial network and resolve it into a multi-dimensional control signal, send the multi-dimensional control signal to the electromagnetic coil array, combine with the induced current, stimulate the directional flow and nonlinear oscillation of the molten pool according to the resonance effect, obtain the oscillation parameters and predict the corresponding thermodynamic behavior of the molten pool according to the oscillation parameters, feed back the thermodynamic prediction results to the central control unit, and optimize and produce the laser parameters and electromagnetic field; The third unit is used to detect the surface morphology of parts in real time according to the visual sensor, extract the defect characteristic parameter vector and add it to the pre-trained physical driving model, invert the defect influencing factors in combination with the digital twin model, obtain the defect cause probability distribution and obtain the defect risk level in combination with the dynamic evolution trend prediction, take the defect cause probability distribution, the defect characteristic parameter vector and the defect risk level as input, generate a dynamic control curve and discretize it into a multi-dimensional control instruction vector, update the electromagnetic field parameters and laser parameters based on the multi-dimensional control instruction vector and the pre-set high-speed feedback channel, generate initial control model parameters in combination with the manufacturing task and material type through the meta-learning algorithm, update the pre-set control model according to the dynamic control curve, output the optimized control model parameters and perform crack suppression operation again.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images