Electromagnetic field composite powder feeding type plasma arc additive manufacturing formability regulation and control method and system

Abstract

The invention provides an electromagnetic field composite powder feeding type plasma arc additive manufacturing formability regulation and control method and system, and relates to the technical field of plasma arc additive manufacturing. The flowing direction and speed of liquid metal in a molten pool in the deposition process are regulated and controlled through the target composite electromagnetic field, and the morphology and the internal organization structure of a deposition layer are controlled; in the process of regulating and controlling the liquid metal in the molten pool, process parameters and molten pool dynamic image data in the deposition process are collected in real time, and a pre-trained state recognition model is used for conducting self-adaptive adjustment on the process parameters of a subsequent deposition layer; and after deposition of each layer is completed, a laser scanning system is adopted for measuring the actual geometric size and surface appearance of the deposition layer, deposition manufacturing of the next layer is executed based on the optimized composite electromagnetic field control parameters, the corrected layered slicing data and the process path, and the execution is executed circularly till manufacturing of the whole workpiece is completed.

Description

Technical Field

[0001] The present invention relates to plasma arc additive manufacturing technology, and in particular to an electromagnetic field composite powder feeding type plasma arc additive manufacturing formability control method and system. Background Art

[0002] Traditional plasma arc additive manufacturing methods have limitations in processing complex geometries and controlling material properties. The main problems include insufficient geometric accuracy, poor surface quality, and many internal defects. These problems mainly stem from insufficient control over the behavior of the molten pool and the lack of an effective real-time monitoring and adjustment mechanism. In the prior art, although there have been attempts to use magnetic fields to control the molten pool, most of them are limited to a single magnetic field, making it difficult to achieve precise local control. This single magnetic field method is inadequate when processing complex shapes or large-sized workpieces, and cannot meet the differentiated control requirements of the molten pool behavior in different regions. In addition, existing methods often lack an effective closed-loop control system, making it difficult to cope with various changes and interferences in the manufacturing process, such as fluctuations in material properties and changes in ambient temperature.

[0003] Therefore, it is of great significance to develop a new additive manufacturing method that can accurately control the melt pool behavior and achieve adaptive adjustment. Summary of the invention

[0004] The embodiments of the present invention provide a method and system for controlling the formability of electromagnetic field composite powder feeding plasma arc additive manufacturing, which can solve the problems in the prior art.

[0005] According to a first aspect of the embodiments of the present invention, Provided is an electromagnetic field composite powder feeding plasma arc additive manufacturing formability control method, comprising: Acquire the three-dimensional model data of the workpiece to be manufactured and perform layered slicing processing to generate layered slicing data, plan the process path of plasma arc additive manufacturing based on the layered slicing data, control the plasma arc welding gun to move along a preset trajectory according to the process path, and at the same time, convey metal powder to the molten pool area through a powder feeding device, apply a composite electromagnetic field in the molten pool area, the composite electromagnetic field is generated by a plurality of independently controlled electromagnetic coils, and according to the geometric characteristics of the current manufacturing layer, dynamically adjust the current parameters of each electromagnetic coil to form a target composite electromagnetic field, and regulate the flow direction and speed of the liquid metal in the molten pool during the deposition process through the target composite electromagnetic field to control the morphology and internal organizational structure of the deposited layer; In the process of regulating the liquid metal in the molten pool, the process parameters and the dynamic image data of the molten pool in the deposition process are collected in real time, and the process parameters and the dynamic image data of the molten pool are analyzed in real time by using a pre-trained state recognition model to identify abnormal deposition states, and a dynamic mapping relationship between the process parameters and the deposition quality is established based on the analysis results, and the process parameters of the subsequent deposition layers are adaptively adjusted according to the dynamic mapping relationship; After completing the deposition of each layer, a laser scanning system is used to measure the actual geometric dimensions and surface morphology of the deposited layer to obtain measurement data, which are compared with the theoretical model generated by preset layered slicing data to calculate the geometric error and surface roughness. Based on the geometric error and surface roughness, the layered slicing data and process path of the next deposition layer are corrected, the cumulative error trend of the geometric error is analyzed, and according to the cumulative error trend, the control parameters of the composite electromagnetic field are optimized. Based on the optimized composite electromagnetic field control parameters, the corrected layered slicing data and process path, the deposition manufacturing of the next layer is executed, and the cycle is performed until the manufacturing of the entire workpiece is completed.

[0006] In an optional embodiment, Acquiring three-dimensional model data of a workpiece to be manufactured and performing layered slicing processing to generate layered slicing data, planning a process path for plasma arc additive manufacturing based on the layered slicing data, and controlling a plasma arc welding gun to move along a preset trajectory according to the process path includes: Acquire three-dimensional model data of a workpiece to be manufactured, convert the three-dimensional model data into a multi-resolution voxel representation by using an octree algorithm, dynamically adjust the voxel size by calculating a local curvature tensor to obtain a voxelized model, perform adaptive layering processing on the voxelized model, optimize the layer thickness by using a gradient descent method, and generate layered data; Based on the layered data, a phase method topology optimization algorithm is introduced, and the stress distribution of the workpiece is predicted by combining finite element analysis to generate optimized filling structure data. A multi-objective particle swarm optimization algorithm is used to plan the process path of the filling structure data, and a multi-objective function including forming accuracy, surface quality, residual stress and manufacturing efficiency is established to obtain an initial process path. The initial process path is optimized by using a Markov decision process, the path direction between adjacent layers is designed, a three-dimensional balance of residual stress is achieved, and a tree-like support structure is automatically generated for the cantilever structure to obtain an optimized process path; Based on the optimized process path, an inverse kinematics algorithm is used to generate a six-degree-of-freedom motion trajectory instruction of a plasma arc welding gun, and a Bezier curve is introduced to smooth the motion trajectory. In the motion trajectory instruction, a powder feeding amount and arc parameter dynamic adjustment instruction based on fuzzy logic control are embedded to form a comprehensive process instruction. A distributed computing framework is used to convert the comprehensive process instructions into control instructions that can be recognized by the equipment, and deposition process data is collected in real time. The deposition process data is fused using a Kalman filter algorithm to obtain fused real-time status data. According to the control instructions and the fused real-time status data, a model predictive control strategy is adopted to dynamically adjust the motion trajectory and process parameters of the plasma arc welding gun, and the plasma arc welding gun is controlled to move along the motion trajectory through an adaptive control algorithm.

[0007] In an optional embodiment, According to the control instructions and the fused real-time status data, the motion trajectory and process parameters of the plasma arc welding gun are dynamically adjusted by using a model predictive control strategy, and the plasma arc welding gun is controlled to move along the motion trajectory by using an adaptive control algorithm, including: A nonlinear state space model of plasma arc welding process is established, which includes state transfer equations and output equations. The state variables include welding gun position, posture and speed, the control variables include joint driving torque and process parameters, and the output variables include weld geometry characteristics and molten pool state. Based on the nonlinear state space model, a nonlinear model predictive control framework is constructed, and an objective function including an output tracking error and a control increment is defined; A real-time iterative method is used to solve the optimization problem of model predictive control. In each control cycle, model linearization, quadratic programming problem construction and iterative solution are performed to obtain a set of optimized control sequences. Design a model reference adaptive control algorithm, establish a reference model, calculate the adaptive law based on the tracking error between the actual system output and the reference model output, and dynamically adjust the controller parameters; The extended Kalman filter is used to fuse multi-source data from visual sensors, force sensors, and thermocouples, and the prior state estimate is calculated based on the current state estimate and control input. The posterior state estimate is updated using the measured data to obtain the optimal estimate of the system state. The optimal estimate of the system state is input into a nonlinear model predictive controller, an optimized control sequence is calculated, a comprehensive control instruction is generated based on the optimized control sequence and the compensation amount of the adaptive control algorithm, the comprehensive control instruction is applied to the drive system of the plasma arc welding gun, the motion trajectory and process parameters of the welding gun are adjusted, and the plasma arc welding gun is controlled to move according to the adjusted motion trajectory and process parameters.

[0008] In an optional embodiment, Using a pre-trained state recognition model, the process parameters and the molten pool dynamic image data are analyzed in real time to identify abnormal deposition states including: The multi-sensor fusion system is used to collect the process parameters and molten pool dynamic image data in the deposition process in real time, and the image preprocessing algorithm is used to perform real-time denoising and enhancement on the molten pool dynamic image data to obtain preprocessed molten pool dynamic image data, and the process parameters are normalized to obtain normalized process parameter data; The pre-processed molten pool dynamic image data and the normalized process parameter data are time synchronized and fused to obtain multi-modal input data, and the multi-modal input data is input into a pre-trained multi-label classification network to obtain an abnormal state prediction result; Based on the abnormal state prediction result, a gradient weighted class activation mapping algorithm is used to generate a class activation heat map, the class activation heat map is superimposed with the preprocessed molten pool dynamic image data, and an adaptive threshold segmentation algorithm is used to locate the abnormal area boundary to obtain the abnormal area segmentation result; Calculate the initial confidence score of each abnormal state prediction result, and use random dropout technology to perform multiple forward propagations on the multi-label classification network to calculate the variance of each abnormal state prediction result to obtain the uncertainty score of the model prediction; Based on the initial confidence score and the uncertainty score, a comprehensive confidence score of each abnormal state is calculated, and a time series analysis is performed on the comprehensive confidence scores of the abnormal states of multiple consecutive frames of images to obtain a final confidence score after time series correction; According to the abnormal region segmentation result and the final confidence score after the timing correction, combined with a pre-established abnormal state feature library, the abnormal deposition state, position, area and severity in the plasma arc additive manufacturing process are determined.

[0009] In an optional embodiment, Based on the initial confidence score and the uncertainty score, the formula for calculating the comprehensive confidence score of each abnormal state is as follows: ;

[0010] Among them, S i，t represents the comprehensive confidence score of abnormal state i at time t, t represents the current time, t0 represents the starting time of the image sequence, β i represents the weight parameter of state i, represents the initial confidence score of abnormal state i, λ represents the uncertainty control coefficient, represents the uncertainty score of abnormal state i, θ represents the uncertainty attenuation coefficient, S i，t-1 represents the comprehensive confidence score of abnormal state i at time t-1, Represents the set of uncertainty scores for all abnormal states.

[0011] In an optional embodiment, The actual geometric dimensions and surface morphology of the deposited layer are measured by a laser scanning system to obtain measurement data, and the measurement data are compared with the theoretical model generated by the preset layered slicing data to calculate the geometric error and surface roughness. Based on the geometric error and surface roughness, the layered slicing data and process path of the next deposited layer are corrected, and the cumulative error trend of the geometric error is analyzed. According to the cumulative error trend, the control parameters of the composite electromagnetic field are optimized, including: A laser scanning system is used to measure the surface of the current layer after deposition to obtain three-dimensional morphology information, the three-dimensional morphology information is compared with a theoretical model generated based on a computer-aided design model and preset layered slice data, and an iterative closest point algorithm is used to perform three-dimensional registration to obtain an actual model after registration; Based on the actual model after registration, the geometric error and surface roughness of the current layer are calculated, where the geometric error includes contour error and thickness error, and the surface roughness is represented by the arithmetic mean roughness value; According to the geometric error, an adaptive algorithm is used to correct the layered slice data of the next deposition layer, the slice profile is adjusted by a reverse offset technology for the profile error, and the thickness error is corrected by dynamically adjusting the interlayer height to obtain the corrected layered slice data; Based on the surface roughness, the process path of the next deposition layer is corrected, and the scanning path and the overlap rate are dynamically adjusted by using an adaptive scanning strategy to obtain a corrected process path; The cumulative trend of the geometric error is analyzed, and a prediction model of the influence of thermal deformation and residual stress on geometric accuracy is established by using a long short-term memory neural network. Based on the prediction model, the influence of the electromagnetic field on the molten pool behavior is simulated by a multi-physical field coupling model, and the molten pool morphology and temperature distribution under different electromagnetic field parameters are predicted to obtain the prediction result of the molten pool behavior; Based on the obtained molten pool behavior prediction results, the static magnetic field intensity, alternating electromagnetic field frequency and intensity are set as optimization variables, and chromosome encoding is constructed. The cumulative geometric error under different parameter combinations is calculated based on the prediction model as the fitness function, and a new generation of population is generated through selection, crossover and mutation operations. The process of evaluating fitness, selecting and generating a new generation of population is repeated until the preset number of iterations is reached, and the individual with the best fitness is selected and decoded to obtain the optimized composite electromagnetic field control parameters.

[0012] In an optional embodiment, A prediction model of the influence of thermal deformation and residual stress on geometric accuracy is established by using a long short-term memory neural network. Based on the prediction model, the influence of the electromagnetic field on the molten pool behavior is simulated by a multi-physics field coupling model, and the molten pool morphology and temperature distribution under different electromagnetic field parameters are predicted. The prediction results of the molten pool behavior include: Based on the cumulative trend of geometric errors, a long short-term memory neural network model is constructed, in which the input layer receives the process parameters, temperature field data and geometric error data of the historical layer, and the output layer predicts the geometric error of the next layer; Using mean square error as the loss function, the constructed long short-term memory neural network model is trained by the optimizer to obtain a trained geometric error prediction model; Constructing a multi-physics field coupling model including electromagnetic field, thermal field and flow field, and setting initial conditions and boundary conditions, obtaining current density and magnetic field distribution by solving electromagnetic field equations in the multi-physics field coupling model, and calculating Joule heat and electromagnetic force; Substitute the obtained Joule heat into the thermal field equation to solve the temperature field distribution, substitute the obtained electromagnetic force into the flow field equation to solve the velocity field and pressure field, update the molten pool morphology and free surface based on the obtained temperature field distribution, velocity field and pressure field, and use the volume fluid method to track the molten pool interface; Repeat the solution of the electromagnetic field, thermal field and flow field equations until the preset time step is reached to obtain the prediction results of the molten pool morphology and temperature distribution. Then change the electromagnetic field parameters and repeat the solution steps of the multi-physics field coupling model to obtain the molten pool morphology and temperature distribution under different electromagnetic field parameters, and finally obtain the prediction results of the molten pool behavior.

[0013] According to a second aspect of the embodiments of the present invention, Provided is an electromagnetic field composite powder feeding plasma arc additive manufacturing formability control system, comprising: The first unit is used to obtain the three-dimensional model data of the workpiece to be manufactured and perform layered slicing processing to generate layered slicing data, plan the process path of plasma arc additive manufacturing based on the layered slicing data, control the plasma arc welding gun to move along a preset trajectory according to the process path, and at the same time, convey metal powder to the molten pool area through a powder feeding device, and apply a composite electromagnetic field in the molten pool area, wherein the composite electromagnetic field is generated by a plurality of independently controlled electromagnetic coils, and the current parameters of each electromagnetic coil are dynamically adjusted according to the geometric characteristics of the current manufacturing layer to form a target composite electromagnetic field, and the flow direction and speed of the liquid metal in the molten pool during the deposition process are regulated by the target composite electromagnetic field to control the morphology and internal organizational structure of the deposited layer; The second unit is used to collect process parameters and dynamic image data of the molten pool in the deposition process in real time during the process of regulating the liquid metal in the molten pool, and to analyze the process parameters and the dynamic image data of the molten pool in real time by using a pre-trained state recognition model to identify abnormal deposition states, and to establish a dynamic mapping relationship between process parameters and deposition quality based on the analysis results, and to adaptively adjust the process parameters of subsequent deposition layers according to the dynamic mapping relationship; The third unit is used to measure the actual geometric size and surface morphology of the deposited layer by using a laser scanning system after completing the deposition of each layer, obtain measurement data, compare the measurement data with the theoretical model generated by preset layered slicing data, calculate the geometric error and surface roughness, correct the layered slicing data and process path of the next deposited layer based on the geometric error and surface roughness, analyze the cumulative error trend of the geometric error, optimize the control parameters of the composite electromagnetic field according to the cumulative error trend, execute the deposition manufacturing of the next layer based on the optimized composite electromagnetic field control parameters and the corrected layered slicing data and process path, and execute the cycle until the manufacturing of the entire workpiece is completed.

[0014] 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.

[0015] A fourth aspect of the embodiments of the present invention is: 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.

[0016] In this embodiment, through real-time monitoring and adaptive adjustment of the system, geometric deviations are effectively reduced, making the manufacturing of complex structures more accurate. The composite electromagnetic field used in the method accurately controls the behavior of the molten pool, greatly improves the surface quality of the workpiece, reduces surface defects, and reduces surface roughness, thereby reducing the need for subsequent processing and improving production efficiency. Through local differentiated control, a gradient distribution of performance is achieved within a single workpiece, which opens up new possibilities for the design and manufacture of composite components with specific functions. At the same time, the method significantly improves process stability, and through real-time identification and correction of abnormal conditions by deep learning algorithms, defects in the manufacturing process are greatly reduced, and product consistency and reliability are improved. By adjusting process parameters in real time, the method can improve production efficiency, reduce material waste, and reduce production costs while ensuring quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a schematic diagram of the process of the electromagnetic field composite powder feeding plasma arc additive manufacturing formability control method according to an embodiment of the present invention; Figure 2 It is a structural schematic diagram of the electromagnetic field composite powder feeding plasma arc additive manufacturing formability control system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0018] 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.

[0019] 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.

[0020] Figure 1 FIG. 1 is a flow chart of a method for controlling the formability of electromagnetic field composite powder feeding plasma arc additive manufacturing according to an embodiment of the present invention. Figure 1 As shown, the method includes: S101. Acquire the three-dimensional model data of the workpiece to be manufactured and perform layered slicing processing to generate layered slicing data, plan the process path of plasma arc additive manufacturing based on the layered slicing data, control the plasma arc welding gun to move along a preset trajectory according to the process path, and at the same time, convey metal powder to the molten pool area through a powder feeding device, and apply a composite electromagnetic field in the molten pool area. The composite electromagnetic field is generated by a plurality of independently controlled electromagnetic coils. According to the geometric characteristics of the current manufacturing layer, the current parameters of each electromagnetic coil are dynamically adjusted to form a target composite electromagnetic field. The target composite electromagnetic field is used to regulate the flow direction and speed of the liquid metal in the molten pool during the deposition process, so as to control the morphology and internal structure of the deposited layer.

[0021] First, the 3D model data of the workpiece to be manufactured is obtained and sliced. This step is completed by computer-aided design software, which converts the 3D model of the workpiece into a series of 2D slice data. The thickness of each slice can be adjusted according to the process requirements, usually between 0.1mm and 1mm. For example, for a cylindrical workpiece with a height of 100mm, it can be divided into 500 slices with a thickness of 0.2mm.

[0022] Based on the generated layered slice data, the process path of plasma arc additive manufacturing is planned. Path planning takes into account multiple factors, including slice shape, filling method, and weld overlap rate. For slices of complex shapes, adaptive path planning algorithms such as equidistant offset method or curvature adaptive method can be used to ensure uniform material deposition. Taking the above-mentioned cylinder as an example, concentric circle filling method can be used, spiral filling from the outer circle to the inside, and the weld overlap rate is set to 30%.

[0023] According to the planned process path, the plasma arc welding gun is controlled to move along the preset trajectory. Parameters such as the moving speed, swing amplitude and frequency of the welding gun need to be precisely controlled according to the process requirements. At the same time, the metal powder is delivered to the molten pool area through the powder feeding device. The control of the powder feeding amount is crucial to the deposition quality, and the closed-loop control system can be used to adjust the powder feeding amount in real time. For example, for titanium alloy powder, the initial powder feeding amount can be set to 2g / min, and the welding gun moving speed is 5mm / s.

[0024] A composite electromagnetic field is applied to the molten pool area, which is generated by multiple independently controlled electromagnetic coils. These electromagnetic coils can be arranged around the welding gun or under the workbench, usually 3 to 8 coils are used. The position, size and winding method of each coil are optimized to achieve precise control of the molten pool. For example, 4 water-cooled coils with a diameter of 50mm and 200 turns can be used, evenly distributed around the welding gun.

[0025] According to the geometric features of the current manufacturing layer, the current parameters of each electromagnetic coil are dynamically adjusted to form a target composite electromagnetic field. This requires the establishment of a mapping relationship between geometric features and electromagnetic field distribution. The pre-trained neural network model can be used to quickly calculate the optimal current parameter combination based on the shape characteristics of the current slice. For example, for a circular cross-section, a more uniform radial electromagnetic field may be required; for a rectangular cross-section, a stronger electromagnetic field may need to be applied in the long side direction to suppress the expansion of the molten pool.

[0026] The target composite electromagnetic field is used to control the flow direction and speed of the liquid metal in the molten pool during the deposition process, thereby controlling the morphology and internal structure of the deposited layer. Under the action of the electromagnetic field, electromagnetic force and electromagnetic stirring effects will be generated in the molten pool, which can change the flow pattern of the liquid metal and affect the solidification process. By precisely controlling the intensity and distribution of the electromagnetic field, the shape of the molten pool, temperature distribution and solidification structure can be controlled.

[0027] In an optional embodiment, Acquiring three-dimensional model data of a workpiece to be manufactured and performing layered slicing processing to generate layered slicing data, planning a process path for plasma arc additive manufacturing based on the layered slicing data, and controlling a plasma arc welding gun to move along a preset trajectory according to the process path includes: Acquire three-dimensional model data of a workpiece to be manufactured, convert the three-dimensional model data into a multi-resolution voxel representation by using an octree algorithm, dynamically adjust the voxel size by calculating a local curvature tensor to obtain a voxelized model, perform adaptive layering processing on the voxelized model, optimize the layer thickness by using a gradient descent method, and generate layered data; Based on the layered data, a phase method topology optimization algorithm is introduced, and the stress distribution of the workpiece is predicted by combining finite element analysis to generate optimized filling structure data. A multi-objective particle swarm optimization algorithm is used to plan the process path of the filling structure data, and a multi-objective function including forming accuracy, surface quality, residual stress and manufacturing efficiency is established to obtain an initial process path. The initial process path is optimized by using a Markov decision process, the path direction between adjacent layers is designed, a three-dimensional balance of residual stress is achieved, and a tree-like support structure is automatically generated for the cantilever structure to obtain an optimized process path; Based on the optimized process path, an inverse kinematics algorithm is used to generate a six-degree-of-freedom motion trajectory instruction of a plasma arc welding gun, and a Bezier curve is introduced to smooth the motion trajectory. In the motion trajectory instruction, a powder feeding amount and arc parameter dynamic adjustment instruction based on fuzzy logic control are embedded to form a comprehensive process instruction. A distributed computing framework is used to convert the comprehensive process instructions into control instructions that can be recognized by the equipment, and deposition process data is collected in real time. The deposition process data is fused using a Kalman filter algorithm to obtain fused real-time status data. According to the control instructions and the fused real-time status data, a model predictive control strategy is adopted to dynamically adjust the motion trajectory and process parameters of the plasma arc welding gun, and the plasma arc welding gun is controlled to move along the motion trajectory through an adaptive control algorithm.

[0028] Exemplarily, first, the 3D model data of the workpiece to be manufactured is obtained. This can be done by computer-aided design software or 3D scanning equipment. For example, for a complex turbine blade model, its 3D data may contain millions of triangular facets.

[0029] The octree algorithm is used to convert the 3D model data into a multi-resolution voxel representation. The octree algorithm recursively divides the space and divides the model into cubic voxels of different sizes. Initially, the maximum voxel size can be set to 5mm and the minimum voxel size can be set to 0.1mm. For a turbine blade model, 10-15 layers of recursive division may be required to achieve the desired accuracy.

[0030] The voxel size is dynamically adjusted by calculating the local curvature tensor. In high curvature areas of the model, such as the leading and trailing edges of the blade, smaller voxels are used to preserve geometric details. In flat areas, such as the middle of the blade, larger voxels are used to reduce the amount of data. The curvature threshold can be set to 0.5mm -1 , when the local curvature exceeds this value, the voxel is further subdivided.

[0031] The voxelized model is adaptively layered. Considering the characteristics of plasma arc additive manufacturing, the initial layer thickness can be set to 0.3mm. Then the gradient descent method is used to optimize the layer thickness to minimize the number of layers while ensuring geometric accuracy. During the optimization process, the geometric error tolerance can be set to 0.05mm, the maximum allowable layer thickness is 0.5mm, and the minimum layer thickness is 0.2mm. After optimization, the turbine blade model may be divided into 200-300 layers.

[0032] Based on the layered data, the phase method topology optimization algorithm is introduced, combined with finite element analysis to predict the stress distribution of the workpiece. First, a finite element model of the workpiece is established, taking into account the thermal stress and mechanical load during the manufacturing process. Then the material distribution is iteratively optimized through the phase method, with the goal of minimizing the material usage while meeting the strength requirements. For turbine blades, 50-100 iterations may be required to obtain the optimal internal structure.

[0033] The multi-objective particle swarm optimization algorithm is used to plan the process path for the filling structure data. A multi-objective function including forming accuracy, surface quality, residual stress and manufacturing efficiency is established. The particle swarm size can be set to 100 and the maximum number of iterations can be set to 200. During the optimization process, the weight of forming accuracy can be set to 0.4, surface quality to 0.3, residual stress to 0.2, and manufacturing efficiency to 0.1. Through iterative optimization, a set of Pareto optimal process path solutions are obtained.

[0034] The initial process path is optimized using the Markov decision process. The path direction between adjacent layers is designed to achieve a three-dimensional balance of residual stress. The path direction can be discretized into four options: 0°, 45°, 90°, and 135°. The optimal direction combination is selected through a reinforcement learning algorithm. For each layer, the state of the current layer and the two adjacent layers is considered, the reward value is calculated, and the strategy is updated. After 1000-2000 iterations, the path direction sequence with the minimum residual stress can be obtained.

[0035] Automatically generate a tree-like support structure for the cantilever structure. First, identify the cantilever area that needs support, and then use a fractal algorithm to generate the tree-like support. The trunk diameter of the support structure can be set to 2mm, and the branch diameter decreases step by step, with a minimum diameter of 0.5mm. The support density decreases with increasing height, and the density at the top is about 1 / 3 of that at the bottom.

[0036] Based on the optimized process path, the inverse kinematics algorithm is used to generate the six-degree-of-freedom motion trajectory instructions of the plasma arc welding gun. Considering the mechanical structure and workspace limitations of the welding gun, the path points in the Cartesian coordinate system are converted into an angle sequence in the joint space. The Jacobian matrix method can be used for solution, and the iteration accuracy is set to 0.01mm.

[0037] Introduce Bezier curve to smooth the motion trajectory. Select the third-order Bezier curve and insert control points at the turning points of the path to ensure the continuity and smoothness of the trajectory. The selection of control points needs to consider the acceleration limit of the welding gun. The maximum acceleration is usually set to 2m / s 2 .

[0038] Embed the dynamic adjustment instructions of powder feeding amount and arc parameters based on fuzzy logic control in the motion trajectory instructions. Establish a fuzzy rule base including molten pool temperature, weld width and height. The membership function of the input variables can be divided into three levels: low, medium and high. The output variables include powder feeding amount adjustment, arc current adjustment and arc voltage adjustment. Dynamic adjustment of parameters is achieved through 50-100 fuzzy rules.

[0039] A distributed computing framework is used to convert comprehensive process instructions into control instructions that can be recognized by the equipment. Distributed computing platforms such as Apache Spark can be used to distribute instruction parsing and conversion tasks to multiple computing nodes for parallel processing. For large workpieces, millions of instructions may need to be processed, and distributed processing can reduce computing time from hours to minutes.

[0040] Real-time collection of deposition process data, including molten pool images, temperature field distribution, arc parameters, etc. A high-speed camera (1000fps) can be used to collect molten pool images, an infrared thermal imager (60Hz) to measure the temperature field, and high-frequency sampling (10kHz) to record arc voltage and current.

[0041] The deposition process data is fused using the Kalman filter algorithm. A state space model including the geometric characteristics of the molten pool, temperature distribution, and arc parameters is established. The measurement noise covariance can be determined by the sensor accuracy, while the process noise covariance needs to be calibrated experimentally. The filter update frequency is set to 100 Hz to balance real-time performance and computational load.

[0042] According to the control instructions and the fused real-time status data, the model predictive control strategy is used to dynamically adjust the motion trajectory and process parameters of the plasma arc welding gun. A multi-physics field coupling model including heat conduction, fluid dynamics and solidification dynamics is established. The prediction time domain can be set to 0.5s and the control time domain is 0.1s. In each control cycle, the optimal control input sequence is obtained by solving the optimization problem.

[0043] The plasma arc welding gun is controlled to move along the motion trajectory through an adaptive control algorithm. A combined control strategy is adopted, including feedforward control and feedback control. Feedforward control calculates the ideal control input based on the kinematic model, while feedback control compensates according to the actual position error. The controller parameters are adjusted in real time through an online identification algorithm to adapt to different working conditions.

[0044] In this embodiment, by introducing technologies such as multi-resolution voxel representation, adaptive stratification, intelligent path planning and real-time control, the accuracy and efficiency of plasma arc additive manufacturing are significantly improved. This method can effectively handle complex geometric shapes, optimize internal structures and reduce material waste. Through multi-objective optimization and Markov decision process, a comprehensive balance of forming quality, efficiency and residual stress is achieved. Real-time data fusion and model predictive control ensure the stability and consistency of the manufacturing process. In addition, the method has good adaptability and scalability, and can be flexibly adjusted according to different materials and process requirements. Finally, this method can significantly improve the technical level of plasma arc additive manufacturing and provide strong support for high-end manufacturing fields such as aerospace and energy.

[0045] In an optional embodiment, According to the control instructions and the fused real-time status data, the motion trajectory and process parameters of the plasma arc welding gun are dynamically adjusted by using a model predictive control strategy, and the plasma arc welding gun is controlled to move along the motion trajectory by using an adaptive control algorithm, including: A nonlinear state space model of plasma arc welding process is established, which includes state transfer equations and output equations. The state variables include welding gun position, posture and speed, the control variables include joint driving torque and process parameters, and the output variables include weld geometry characteristics and molten pool state. Based on the nonlinear state space model, a nonlinear model predictive control framework is constructed, and an objective function including an output tracking error and a control increment is defined; A real-time iterative method is used to solve the optimization problem of model predictive control. In each control cycle, model linearization, quadratic programming problem construction and iterative solution are performed to obtain a set of optimized control sequences. Design a model reference adaptive control algorithm, establish a reference model, calculate the adaptive law based on the tracking error between the actual system output and the reference model output, and dynamically adjust the controller parameters; The extended Kalman filter is used to fuse multi-source data from visual sensors, force sensors, and thermocouples, and the prior state estimate is calculated based on the current state estimate and control input. The posterior state estimate is updated using the measured data to obtain the optimal estimate of the system state. The optimal estimate of the system state is input into a nonlinear model predictive controller, an optimized control sequence is calculated, a comprehensive control instruction is generated based on the optimized control sequence and the compensation amount of the adaptive control algorithm, the comprehensive control instruction is applied to the drive system of the plasma arc welding gun, the motion trajectory and process parameters of the welding gun are adjusted, and the plasma arc welding gun is controlled to move according to the adjusted motion trajectory and process parameters.

[0046] Exemplarily, first, a nonlinear state space model of the plasma arc welding process is established. The model includes state transfer equations and output equations. The state variables include the position, posture and speed of the welding gun, which can be represented by a six-dimensional vector, for example, in which the first three dimensions represent the position of the end of the welding gun in the Cartesian coordinate system, and the last three dimensions represent the posture with Euler angles. The control variables include joint driving torques and process parameters, for example, which can be represented by a nine-dimensional vector, in which the first six dimensions correspond to the driving torques of the six joints, and the last three dimensions represent the current, voltage and wire feeding speed respectively. The output variables include weld geometry and molten pool state, which can be represented by a five-dimensional vector, including weld width, height, depth, and length and width of the molten pool.

[0047] The state transfer equation describes the evolution of the system state over time. Considering the complexity of the plasma arc welding process, a neural network can be used to fit this nonlinear function. For example, a three-layer feedforward neural network can be used, where the input layer contains the current state and control variables, the hidden layer is set with 100 neurons, and the output layer corresponds to the state at the next moment. This neural network is trained with a large amount of experimental data so that it can accurately predict the dynamic characteristics of the system.

[0048] The output equation describes the relationship between the system output and the current state and control variables. A neural network can also be used to model the system. The input layer contains the current state and control variables, the hidden layer has 50 neurons, and the output layer corresponds to the weld geometry and the molten pool state. This neural network improves the prediction accuracy by combining offline training and online fine-tuning.

[0049] Based on the above nonlinear state space model, a nonlinear model predictive control framework is constructed. The objective function is defined, which includes two parts: output tracking error and control increment. The output tracking error reflects the gap between the actual output and the expected output, which can be measured by weighted Euclidean distance. The control increment represents the change in the control variable between two adjacent control cycles. This term is introduced to avoid drastic fluctuations in the control input. In practical applications, the prediction time domain can be set to 20 sampling cycles, and the control time domain can be set to 5 sampling cycles. The weight coefficients of each item in the objective function need to be determined through experiments. For example, the weight of the output tracking error can be set to 0.8, and the weight of the control increment can be set to 0.2.

[0050] A real-time iteration method is used to solve the optimization problem of model predictive control. In each control cycle, the nonlinear model is first linearized. The first-order Taylor expansion can be used to approximate the nonlinear function to obtain the linearized state equation and output equation. Then, a quadratic programming problem is constructed based on the linearized model. The objective function can be expressed as a quadratic form of the state variable and the control variable, and the constraints include the state equation, the output equation, and the upper and lower limit constraints of the control variable. Finally, an iterative solution algorithm is used to solve this quadratic programming problem. Efficient algorithms such as the interior point method or the active set method can be selected to obtain an approximate optimal solution within a limited computing time. Usually, setting the maximum number of iterations to 50 and the convergence threshold to 1e-6 can complete the solution process within 10ms.

[0051] Design a model reference adaptive control algorithm to compensate for model uncertainties and external disturbances. First, establish a reference model that describes the ideal dynamic characteristics of the system. A second-order linear system can be selected as the reference model, where the natural frequency and damping ratio are set according to the welding process requirements. Then, the adaptive law is calculated based on the tracking error between the actual system output and the reference model output. The adaptive law determines how to adjust the controller parameters and can be designed using the MIT rule. Specifically, the adjustment speed of the controller parameters is proportional to the gradient of the tracking error, and the proportionality factor (adaptive gain) needs to be selected based on the system characteristics, usually between 0.1 and 1.

[0052] The extended Kalman filter is used to fuse data from multiple sensors to obtain the optimal estimate of the system state. First, the observation equation of the system needs to be established to describe the relationship between the measured values ​​and the system state. The visual sensor provides measurements of the weld geometry and the molten pool state, the force sensor measures the contact force at the end of the welding gun, and the thermocouple measures the workpiece temperature. These measurements can all be expressed as nonlinear functions of the state variables. Then, the prior state estimate is calculated based on the current state estimate and the control input. This step uses the nonlinear state transfer equation established previously to predict the state at the next moment through numerical integration. Next, the actual measurement data is used to update the posterior state estimate. The update process involves the calculation of the Kalman gain, in which the process noise covariance and the measurement noise covariance need to be set. These two parameters can be determined through offline experiments. For example, the process noise covariance can be set to the unit matrix multiplied by 0.01, and the measurement noise covariance is set according to the accuracy of each sensor.

[0053] The optimal estimate of the system state is input into the nonlinear model predictive controller to calculate the optimized control sequence. This step utilizes the model predictive control framework constructed earlier, takes the state estimate as the initial condition, and solves the optimization problem to obtain the optimal control input for the next few control cycles. Then, the optimized control sequence is combined with the compensation of the adaptive control algorithm to generate a comprehensive control command. The compensation of the adaptive control algorithm is calculated based on the tracking error and the adaptive law. It can correct the results of the model predictive control and improve the robustness of the system.

[0054] Finally, the integrated control instructions are applied to the driving system of the plasma arc welding gun to adjust the motion trajectory and process parameters of the welding gun. The control instructions include joint drive torque and process parameters. The joint drive torque is executed by the servo system of the welding robot to achieve precise motion control of the welding gun. The process parameters are adjusted by the plasma arc welding power supply and wire feeding system, including real-time adjustment of current, voltage and wire feeding speed. The control cycle can be set to 10ms to ensure the real-time and smoothness of the control.

[0055] In this embodiment, by introducing advanced technologies such as nonlinear model predictive control, adaptive control and multi-sensor data fusion, the control accuracy and stability of the plasma arc welding process are significantly improved. This method can effectively handle the nonlinear characteristics and uncertainties in the welding process and achieve precise control of the weld quality. Through model predictive control, the system can predict and optimize future control strategies in advance, effectively reducing the occurrence of welding defects. The adaptive control algorithm enhances the system's adaptability to parameter changes and external interference, ensuring the robustness of the welding process. Multi-sensor data fusion technology improves the accuracy of state estimation and provides a reliable basis for control decisions. In addition, this method has good real-time and scalability, and can adapt to different welding processes and equipment conditions.

[0056] S102. In the process of regulating the liquid metal in the molten pool, the process parameters and dynamic image data of the molten pool in the deposition process are collected in real time, and a pre-trained state recognition model is used to analyze the process parameters and the dynamic image data of the molten pool in real time to identify abnormal deposition states. Based on the analysis results, a dynamic mapping relationship between the process parameters and the deposition quality is established, and according to the dynamic mapping relationship, the process parameters of the subsequent deposition layers are adaptively adjusted.

[0057] In the process of regulating the liquid metal in the molten pool, the process parameters and dynamic image data of the molten pool in the deposition process are collected in real time. The process parameters include arc voltage, current, welding gun movement speed, powder feeding amount, etc., which can be collected in real time through various sensors. The dynamic image of the molten pool is obtained by a high-speed camera with a collection frequency of up to 1000 frames / second to capture the transient changes of the molten pool.

[0058] Using the pre-trained state recognition model, the collected process parameters and dynamic image data of the molten pool are analyzed in real time to identify abnormal deposition states. The model can be built based on convolutional neural networks and long short-term memory networks, and can extract features from images and time series data to determine whether the current deposition state is normal. For example, abnormal states such as overheating of the molten pool, insufficient melting, and unstable powder feeding can be identified.

[0059] Based on the analysis results, a dynamic mapping relationship between process parameters and deposition quality is established. This mapping relationship can be established through machine learning algorithms such as support vector regression or random forest. The mapping relationship will be continuously updated and optimized as the manufacturing process progresses to adapt to changes in material properties and environmental conditions.

[0060] Based on the established dynamic mapping relationship, the process parameters of the subsequent deposition layers are adaptively adjusted. For example, if the molten pool temperature is detected to be too high, the arc power can be reduced or the welding gun movement speed can be increased; if insufficient melting is found, the arc power can be increased or the welding gun movement speed can be reduced. This adaptive adjustment ensures the stability and consistency of the deposition process.

[0061] In an optional embodiment, Using a pre-trained state recognition model, the process parameters and the molten pool dynamic image data are analyzed in real time to identify abnormal deposition states including: The multi-sensor fusion system is used to collect the process parameters and molten pool dynamic image data in the deposition process in real time, and the image preprocessing algorithm is used to perform real-time denoising and enhancement on the molten pool dynamic image data to obtain preprocessed molten pool dynamic image data, and the process parameters are normalized to obtain normalized process parameter data; The pre-processed molten pool dynamic image data and the normalized process parameter data are time synchronized and fused to obtain multi-modal input data, and the multi-modal input data is input into a pre-trained multi-label classification network to obtain an abnormal state prediction result; Based on the abnormal state prediction result, a gradient weighted class activation mapping algorithm is used to generate a class activation heat map, the class activation heat map is superimposed with the preprocessed molten pool dynamic image data, and an adaptive threshold segmentation algorithm is used to locate the abnormal area boundary to obtain the abnormal area segmentation result; Calculate the initial confidence score of each abnormal state prediction result, and use random dropout technology to perform multiple forward propagations on the multi-label classification network to calculate the variance of each abnormal state prediction result to obtain the uncertainty score of the model prediction; Based on the initial confidence score and the uncertainty score, a comprehensive confidence score of each abnormal state is calculated, and a time series analysis is performed on the comprehensive confidence scores of the abnormal states of multiple consecutive frames of images to obtain a final confidence score after time series correction; According to the abnormal region segmentation result and the final confidence score after the timing correction, combined with a pre-established abnormal state feature library, the abnormal deposition state, position, area and severity in the plasma arc additive manufacturing process are determined.

[0062] Exemplarily, first, a multi-sensor fusion system is used to collect process parameters and dynamic image data of the molten pool in real time during the deposition process. Process parameters include current, voltage, wire feeding speed, shielding gas flow, etc., which can be collected at a sampling frequency of 1kHz by a high-precision data acquisition card. The dynamic image data of the molten pool is collected by a high-speed camera with a frame rate set to 500fps and a resolution of 1280x1024 pixels. In order to ensure image quality, a bandpass filter can be used to filter out most of the arc light, and a near-infrared light source can be used for auxiliary lighting.

[0063] The collected dynamic image data of the molten pool are preprocessed in real time. First, the median filter algorithm is used to remove the salt and pepper noise in the image, and the filter window size is set to 3x3 pixels. Then the adaptive histogram equalization algorithm is applied to enhance the image contrast. The image is divided into 8x8 grids, and histogram equalization is performed separately in each grid. The grid boundary effect is eliminated by bilinear interpolation. Finally, the image is smoothed using a Gaussian filter, with the kernel size set to 5x5 pixels and the standard deviation set to 1.0. These preprocessing steps can effectively improve the image quality and highlight the boundaries and internal details of the molten pool.

[0064] At the same time, the process parameters are normalized. The minimum-maximum normalization method is used to map each parameter to the range of 0 to 1. Specifically, for the current parameter, 100A can be mapped to 0 and 300A can be mapped to 1; for the voltage parameter, 20V can be mapped to 0 and 40V can be mapped to 1; for the wire feeding speed, 2m / min can be mapped to 0 and 8m / min can be mapped to 1. This normalization process can eliminate the dimensional differences between different parameters, which is conducive to subsequent data fusion and model training.

[0065] The preprocessed molten pool dynamic image data and the normalized process parameter data are time synchronized and fused. Time synchronization uses the timestamp alignment method to correspond each frame of the image to the closest process parameter sampling point. Data fusion uses the early fusion strategy to splice the process parameter data with the image data as an additional channel. Specifically, the normalized process parameters can be copied and expanded into a matrix of the same size as the image, and then spliced ​​with the RGB image in the channel dimension to form a 7-channel multimodal input data.

[0066] The multimodal input data is input into the pre-trained multi-label classification network to obtain the abnormal state prediction results. The network uses ResNet50 as the backbone network, and the number of output nodes of the last fully connected layer is equal to the number of predefined abnormal state categories (for example, 10 abnormal states). The cross entropy loss function is used during network training, and the Adam optimizer is used. The learning rate is set to 0.001, the batch size is 32, and the number of training rounds is 100. In order to improve the generalization ability of the model, data enhancement techniques are used during training, including random flipping, rotation, scaling, and brightness adjustment.

[0067] Based on the abnormal state prediction results, the gradient weighted class activation mapping algorithm is used to generate a class activation heat map. The algorithm calculates the gradient of the predicted category to the convolution feature map to obtain the importance weight of each feature map, and then superimposes the weighted feature maps to obtain a heat map. The heat map shows the image area that contributes most to the prediction result, which helps to locate the abnormal area.

[0068] The class activation heat map is superimposed with the preprocessed dynamic image data of the melt pool, and the adaptive threshold segmentation algorithm is used to locate the boundary of the abnormal area. Specifically, the heat map is first normalized to a grayscale range of 0-255, and then the Otsu adaptive threshold algorithm is used to calculate the optimal segmentation threshold. The areas above the threshold are marked as potential abnormal areas, and morphological operations (such as opening and closing operations) are used to smooth the boundaries and remove small noise areas. Finally, the contour extraction algorithm is used to obtain the precise boundaries of the abnormal area.

[0069] Calculate the initial confidence score of each abnormal state prediction result. The initial confidence score is directly taken from the probability value output by the multi-label classification network. At the same time, the random dropout technique is used to perform multiple forward propagations (for example, 20 times) on the multi-label classification network, calculate the variance of each abnormal state prediction result, and obtain the uncertainty score of the model prediction. The random dropout ratio is set to 0.2, that is, 20% of the neurons are randomly turned off during each forward propagation.

[0070] Based on the initial confidence score and uncertainty score, the comprehensive confidence score of each abnormal state is calculated. The comprehensive confidence score can be obtained by dividing the initial confidence score by the uncertainty score, which can reduce the weight of high uncertainty predictions. The time series analysis of the comprehensive confidence scores of abnormal states of multiple consecutive frames of images (for example, 10 frames) is performed, and the final confidence score after time series correction is obtained using the exponential weighted moving average method. The exponential weighting factor can be set to 0.8 to balance real-time and stability.

[0071] According to the abnormal area segmentation results and the final confidence score after timing correction, combined with the pre-established abnormal state feature library, the abnormal deposition state, position, area and severity in the plasma arc additive manufacturing process are determined. The abnormal state feature library contains feature descriptions of various typical anomalies, such as spheroidization, spatter, and unfusion. By comparing the test results with the descriptions in the feature library, the type of anomaly can be accurately determined. The abnormal position is determined by the coordinates in the segmentation result, the area is obtained by calculating the number of pixels in the segmented area, and the severity is comprehensively evaluated based on the final confidence score and the abnormal area.

[0072] In an optional embodiment, Based on the initial confidence score and the uncertainty score, the formula for calculating the comprehensive confidence score of each abnormal state is as follows: ;

[0073] Among them, S i，t represents the comprehensive confidence score of abnormal state i at time t, t represents the current time, t0 represents the starting time of the image sequence, β i represents the weight parameter of state i, represents the initial confidence score of abnormal state i, λ represents the uncertainty control coefficient, represents the uncertainty score of abnormal state i, θ represents the uncertainty attenuation coefficient, S i，t-1 represents the comprehensive confidence score of abnormal state i at time t-1, Represents the set of uncertainty scores for all abnormal states.

[0074] In this embodiment, real-time and accurate detection of abnormal conditions in the plasma arc additive manufacturing process is achieved through multi-sensor fusion and deep learning technology. The method has the following technical effects: First, multi-sensor fusion technology provides rich process information, combined with image preprocessing and data normalization, significantly improving the quality and consistency of input data. Secondly, the deep learning multi-label classification network can simultaneously identify multiple abnormal conditions, improving the comprehensiveness and efficiency of detection. Thirdly, the gradient weighted class activation mapping and adaptive threshold segmentation algorithm realize the precise positioning of abnormal areas, providing an important basis for subsequent defect repair. In addition, by introducing random deactivation and time series analysis, the reliability and stability of the prediction results are improved. Finally, combined with the pre-established abnormal state feature library, a comprehensive description of the abnormal state, including type, location, area and severity, is achieved, providing detailed guidance for process parameter optimization and quality control.

[0075] S103. After completing the deposition of each layer, a laser scanning system is used to measure the actual geometric dimensions and surface morphology of the deposited layer to obtain measurement data. The measurement data is compared with the theoretical model generated by preset layered slicing data to calculate the geometric error and surface roughness. Based on the geometric error and surface roughness, the layered slicing data and process path of the next deposition layer are corrected. The cumulative error trend of the geometric error is analyzed. According to the cumulative error trend, the control parameters of the composite electromagnetic field are optimized. Based on the optimized composite electromagnetic field control parameters and the corrected layered slicing data and process path, the deposition manufacturing of the next layer is executed. The cycle is executed until the manufacturing of the entire workpiece is completed.

[0076] After each layer of deposition is completed, a laser scanning system is used to measure the actual geometric dimensions and surface morphology of the deposited layer. The resolution of the laser scanning system can usually reach 0.01mm, which can accurately capture the microscopic morphological features of the deposited layer. After the scanning data is processed, a three-dimensional point cloud model of the deposited layer is generated.

[0077] The measured point cloud data is compared with the theoretical model generated by the preset layered slice data to calculate the geometric error and surface roughness. The geometric error includes height error, contour error, etc., which can be calculated through point cloud registration and geometric comparison algorithms. The surface roughness can be evaluated by analyzing the local undulation characteristics of the point cloud data.

[0078] Based on the calculated geometric error and surface roughness, the layered slice data and process path of the next deposition layer are corrected. For example, if a certain area is too low in height, the material deposition amount in that area can be increased in the next layer; if the edge profile deviates, the process path can be adjusted to compensate for the deviation. This adaptive correction ensures effective control of cumulative errors.

[0079] Analyze the cumulative error trend of geometric errors, which can be achieved by performing time series analysis on error data of consecutive layers. The cumulative error trend reflects the existence of systematic errors in the manufacturing process, such as error accumulation caused by factors such as thermal deformation and residual stress.

[0080] According to the cumulative error trend obtained by analysis, the control parameters of the composite electromagnetic field are optimized. For example, if it is found that the workpiece tends to shrink and deform as the number of layers increases, the electromagnetic field distribution can be adjusted to increase the electromagnetic force in a specific area to suppress the shrinkage trend. This optimization process can be achieved through a reinforcement learning algorithm. The system learns the optimal control strategy by continuously trying different combinations of electromagnetic field parameters.

[0081] Based on the optimized composite electromagnetic field control parameters, the corrected layered slicing data and process path, the deposition manufacturing of the next layer is performed. The whole process is executed in a cycle until the manufacturing of the entire workpiece is completed. Through this closed-loop control and continuous optimization, high-precision and high-quality additive manufacturing is achieved.

[0082] In an optional embodiment, The actual geometric dimensions and surface morphology of the deposited layer are measured by a laser scanning system to obtain measurement data, and the measurement data are compared with the theoretical model generated by the preset layered slicing data to calculate the geometric error and surface roughness. Based on the geometric error and surface roughness, the layered slicing data and process path of the next deposited layer are corrected, and the cumulative error trend of the geometric error is analyzed. According to the cumulative error trend, the control parameters of the composite electromagnetic field are optimized, including: A laser scanning system is used to measure the surface of the current layer after deposition to obtain three-dimensional morphology information, the three-dimensional morphology information is compared with a theoretical model generated based on a computer-aided design model and preset layered slice data, and an iterative closest point algorithm is used to perform three-dimensional registration to obtain an actual model after registration; Based on the actual model after registration, the geometric error and surface roughness of the current layer are calculated, where the geometric error includes contour error and thickness error, and the surface roughness is represented by the arithmetic mean roughness value; According to the geometric error, an adaptive algorithm is used to correct the layered slice data of the next deposition layer, the slice profile is adjusted by a reverse offset technology for the profile error, and the thickness error is corrected by dynamically adjusting the interlayer height to obtain the corrected layered slice data; Based on the surface roughness, the process path of the next deposition layer is corrected, and the scanning path and the overlap rate are dynamically adjusted by using an adaptive scanning strategy to obtain a corrected process path; The cumulative trend of the geometric error is analyzed, and a prediction model of the influence of thermal deformation and residual stress on geometric accuracy is established by using a long short-term memory neural network. Based on the prediction model, the influence of the electromagnetic field on the molten pool behavior is simulated by a multi-physical field coupling model, and the molten pool morphology and temperature distribution under different electromagnetic field parameters are predicted to obtain the prediction result of the molten pool behavior; Based on the obtained molten pool behavior prediction results, the static magnetic field intensity, alternating electromagnetic field frequency and intensity are set as optimization variables, and chromosome encoding is constructed. The cumulative geometric error under different parameter combinations is calculated based on the prediction model as the fitness function, and a new generation of population is generated through selection, crossover and mutation operations. The process of evaluating fitness, selecting and generating a new generation of population is repeated until the preset number of iterations is reached, and the individual with the best fitness is selected and decoded to obtain the optimized composite electromagnetic field control parameters.

[0083] For example, first, a laser scanning system is used to measure the surface of the current layer that has been deposited to obtain three-dimensional morphological information. The laser scanning system uses a combination of line laser projection and high-resolution cameras to reconstruct the three-dimensional surface information through the principle of triangulation. The scanning resolution is set to 0.1mm, and the scanning speed is 100mm / s. A 200mm x 200mm area can be scanned within 30 seconds. The obtained three-dimensional point cloud data contains the height information and microscopic morphological features of the surface.

[0084] The acquired 3D topography information was compared with the theoretical model generated based on the computer-aided design model and preset layered slice data. The theoretical model was obtained by slicing the original CAD model according to the preset layer thickness (e.g., 0.3 mm). The iterative closest point algorithm was used for 3D registration, which achieves registration by repeatedly finding the closest point pairs in the two sets of point clouds and minimizing the distance between them. The number of iterations was set to 50, and the convergence threshold was 0.01 mm. The registration process took into account translation, rotation, and scaling transformations to compensate for possible thermal deformation and shrinkage effects.

[0085] Based on the actual model after registration, the geometric error and surface roughness of the current layer are calculated. Geometric errors include contour error and thickness error. Contour error is obtained by comparing the deviation between the actual contour and the theoretical contour, and can be characterized by maximum deviation and average deviation. For example, the maximum contour deviation of a certain area is 0.5mm, and the average deviation is 0.2mm. The thickness error is calculated by comparing the difference between the actual layer thickness and the ideal layer thickness. For example, the actual average layer thickness is 0.32mm, and the error relative to the ideal layer thickness of 0.3mm is 0.02mm. The surface roughness is characterized by the arithmetic mean roughness value, which is obtained by calculating the average of the absolute value of the surface height deviation. The typical value may be in the range of 10-50μm.

[0086] Based on the calculated geometric error, an adaptive algorithm is used to correct the layered slicing data of the next deposition layer. For contour errors, the slice contour is adjusted by reverse offset technology. Specifically, if the outer contour of a certain area is detected to be 0.3mm larger, the contour of the area will be shrunk inward by 0.3mm when slicing the next layer. For thickness errors, corrections are made by dynamically adjusting the interlayer height. For example, if the average thickness of the current layer is 0.32mm, the theoretical thickness of the next layer can be adjusted to 0.28mm to compensate for the cumulative error. This dynamic adjustment ensures precise control of the overall height.

[0087] Based on the measured surface roughness, the process path of the next deposition layer is corrected. The scanning path and overlap rate are dynamically adjusted using an adaptive scanning strategy. For areas with higher roughness (such as Ra>30μm), the local scanning density can be increased, and the scanning line spacing can be reduced from the standard 0.8mm to 0.6mm. At the same time, the scanning direction can be adjusted to make it at a 45° or 90° angle with the previous layer to break the continuity of the surface texture. For edge areas, the number of contour scans can be increased from the standard 1 to 2 to improve surface smoothness.

[0088] The cumulative trend of geometric errors was analyzed, and a prediction model for the influence of thermal deformation and residual stress on geometric accuracy was established using a long short-term memory neural network. The network contains an LSTM layer (64 units) and two fully connected layers (32 and 16 neurons, respectively). The input features include the geometric errors, process parameters, and ambient temperature of the historical layer, and the output is the predicted geometric error. The model was trained using the Adam optimizer with a learning rate of 0.001, a batch size of 32, and 500 training rounds.

[0089] Based on the established prediction model, the influence of electromagnetic field on the behavior of molten pool is simulated by multi-physics field coupling model. The model comprehensively considers physical processes such as electromagnetic field, heat conduction, fluid dynamics and free surface deformation. The Maxwell equations, Navier-Stokes equations and heat conduction equations are solved by finite element method to predict the morphology and temperature distribution of molten pool under different electromagnetic field parameters. The simulation grid adopts an adaptive encryption strategy, and the grid size of the molten pool area is set to 0.1mm, and the far field area can be relaxed to 1mm. The time step is set to 0.1ms to capture the rapidly changing dynamic behavior of the molten pool.

[0090] Based on the prediction results of the molten pool behavior, the static magnetic field intensity, alternating electromagnetic field frequency and intensity are set as optimization variables to construct chromosome encoding. The static magnetic field intensity range is set to 0-100mT, the alternating electromagnetic field frequency range is 0-1000Hz, and the intensity range is 0-50mT. Each variable is encoded with 8 bits of binary to form a 24-bit chromosome. The initial population size is set to 100 and initialized by random generation.

[0091] Based on the prediction model, the cumulative geometric error under different parameter combinations is calculated as the fitness function. The fitness function is defined as the inverse of the geometric error. The smaller the error, the higher the fitness. The excellent individuals are selected by the roulette selection method, and the selection probability is proportional to the fitness. The crossover operation uses single-point crossover, and the crossover probability is set to 0.8. The mutation operation uses bit flipping, and the mutation probability is set to 0.05.

[0092] Repeat the process of evaluating fitness, selecting and generating a new generation of population until the preset number of iterations (such as 100 generations) is reached or the fitness value tends to stabilize. Select the individual with the best fitness and decode it to obtain the optimized composite electromagnetic field control parameters. For example, the optimal solution may be: static magnetic field intensity 75mT, alternating electromagnetic field frequency 450Hz, intensity 35mT. This set of parameters can minimize the cumulative geometric error while ensuring the stability of the molten pool.

[0093] In this embodiment, firstly, the introduction of the laser scanning system realizes the high-precision measurement of the geometric features of the deposited layer, providing a reliable data basis for subsequent optimization. Secondly, the adaptive slicing and path optimization strategy based on the measurement results can effectively compensate for geometric errors and improve surface quality, significantly improving the forming accuracy. Thirdly, the application of long short-term memory neural network enables the system to learn and predict the long-term effects of thermal deformation and residual stress, improving the foresight of control. In addition, the introduction of multi-physics field coupling model deepens the understanding of molten pool behavior and provides a theoretical basis for electromagnetic field parameter optimization. Finally, the electromagnetic field parameter optimization based on genetic algorithm realizes the fine control of the forming process and further improves the stability and consistency of the forming quality. In general, this method significantly improves the accuracy and controllability of plasma arc additive manufacturing through closed-loop feedback and intelligent optimization, and provides strong support for the manufacture of high-performance and high-precision parts.

[0094] In an optional embodiment, A prediction model of the influence of thermal deformation and residual stress on geometric accuracy is established by using a long short-term memory neural network. Based on the prediction model, the influence of the electromagnetic field on the molten pool behavior is simulated by a multi-physics field coupling model, and the molten pool morphology and temperature distribution under different electromagnetic field parameters are predicted. The prediction results of the molten pool behavior include: Based on the cumulative trend of geometric errors, a long short-term memory neural network model is constructed, in which the input layer receives the process parameters, temperature field data and geometric error data of the historical layer, and the output layer predicts the geometric error of the next layer; Using mean square error as the loss function, the constructed long short-term memory neural network model is trained by the optimizer to obtain a trained geometric error prediction model; Constructing a multi-physics field coupling model including electromagnetic field, thermal field and flow field, and setting initial conditions and boundary conditions, obtaining current density and magnetic field distribution by solving electromagnetic field equations in the multi-physics field coupling model, and calculating Joule heat and electromagnetic force; Substitute the obtained Joule heat into the thermal field equation to solve the temperature field distribution, substitute the obtained electromagnetic force into the flow field equation to solve the velocity field and pressure field, update the molten pool morphology and free surface based on the obtained temperature field distribution, velocity field and pressure field, and use the volume fluid method to track the molten pool interface; Repeat the solution of the electromagnetic field, thermal field and flow field equations until the preset time step is reached to obtain the prediction results of the molten pool morphology and temperature distribution. Then change the electromagnetic field parameters and repeat the solution steps of the multi-physics field coupling model to obtain the molten pool morphology and temperature distribution under different electromagnetic field parameters, and finally obtain the prediction results of the molten pool behavior.

[0095] Exemplarily, first, based on the cumulative trend of geometric errors, a long short-term memory neural network model is constructed. The input layer of the model receives the process parameters, temperature field data and geometric error data of the history layer. Specifically, the process parameters include arc current (such as 100-300A), arc voltage (such as 20-30V), wire feeding speed (such as 2-5m / min) and scanning speed (such as 5-10mm / s). The temperature field data includes the maximum temperature of the molten pool (such as 2000-2500K), the edge temperature of the molten pool (such as 1700-1900K) and the temperature gradient of the heat-affected zone (such as 200-500K / mm). The geometric error data includes contour error (such as ±0.2mm) and thickness error (such as ±0.05mm). The output layer predicts the geometric error of the next layer, including the expected contour error and thickness error.

[0096] The network structure design includes an LSTM layer (containing 64 neurons), two fully connected hidden layers (containing 32 and 16 neurons respectively) and an output layer (2 neurons, corresponding to contour error and thickness error respectively). The LSTM layer can capture long-term dependencies and is suitable for processing time series data. The hidden layer uses the ReLU activation function, and the output layer uses the linear activation function.

[0097] The Adam optimizer was used to train the constructed long short-term memory neural network model using mean square error as the loss function. The learning rate was set to 0.001, the batch size was 32, and the number of training rounds was 500. The training data set contained 1000 sets of historical data, of which 80% were used for training and 20% for validation. During the training process, when the loss on the validation set did not decrease for 10 consecutive rounds, the early stopping strategy was used to terminate the training to prevent overfitting. Finally, on the test set, the model's predicted root mean square error for contour error reached 0.03mm, and the predicted root mean square error for thickness error reached 0.01mm, and the trained geometric error prediction model was obtained.

[0098] Next, a multi-physics coupling model including electromagnetic field, thermal field and flow field is constructed. The electromagnetic field model is based on the Maxwell equations and takes into account the interaction between the plasma arc, the molten pool and the workpiece. The thermal field model is based on the heat conduction equation and takes into account factors such as Joule heat, plasma arc heat input, phase change latent heat and surface radiation. The flow field model is based on the Navier-Stokes equations and takes into account driving forces such as surface tension, buoyancy, electromagnetic force and Marangoni effect.

[0099] Set initial conditions and boundary conditions. Initial conditions include ambient temperature (e.g. 300K), initial velocity field (stationary state) and initial pressure field (atmospheric pressure). Boundary conditions include heat exchange coefficients at the bottom and sides of the workpiece (e.g. 50 W / m²·K), surface radiation emissivity (e.g. 0.8) and surface tension temperature coefficient (e.g. -0.0005 N / m·K).

[0100] By solving the electromagnetic field equations in the multi-physics field coupling model, the current density and magnetic field distribution are obtained, and the Joule heat and electromagnetic force are calculated. The finite element method is used for discretization, and the grid size is set to 0.1mm in the molten pool area and can be relaxed to 1mm in the far field area. The time step is set to 0.1ms to capture the rapidly changing electromagnetic field dynamics. Typical current density values ​​may be around 10 6 -10 7 The magnetic field strength is in the range of 0.01-0.1 T.

[0101] The obtained Joule heat is substituted into the thermal field equation to solve the temperature field distribution. The Joule heat is added to the heat conduction equation as a body heat source, and the plasma arc heat input is considered as a surface heat source. The temperature field distribution obtained by the solution includes the molten pool area (temperature higher than the melting point of the material, such as about 1800K for steel), the heat affected zone (temperature between 700-1800K) and the unaffected area.

[0102] Substitute the obtained electromagnetic force into the flow field equation to solve the velocity field and pressure field. The electromagnetic force is added to the momentum equation as a volume force term, and affects the flow of the molten pool together with buoyancy, surface tension gradient force, etc. The typical molten pool flow velocity may be in the range of 0.1-1 m / s, and the pressure field changes in the range of hundreds to thousands of Pa.

[0103] Based on the obtained temperature field distribution, velocity field and pressure field, the molten pool morphology and free surface are updated. The molten pool interface is tracked by the volume fluid method, which describes the distribution of liquid and gas phases by solving the volume fraction transport equation. The interface reconstruction uses the piecewise linear interface calculation (PLIC) method to accurately capture the deformation of the free surface of the molten pool.

[0104] Repeat the solution of electromagnetic field, thermal field and flow field equations until the preset time step (such as 10ms) is reached to obtain the prediction results of molten pool morphology and temperature distribution. Then change the electromagnetic field parameters, such as static magnetic field intensity (0-100mT), alternating electromagnetic field frequency (0-1000Hz) and intensity (0-50mT), repeat the multi-physics field coupling model solution steps, and obtain the molten pool morphology and temperature distribution under different electromagnetic field parameters. By systematically changing these parameters, a comprehensive database of molten pool behavior prediction results is finally obtained.

[0105] In this embodiment, firstly, the long short-term memory neural network model can effectively capture the cumulative trend of geometric errors, and improve the accuracy and stability of the prediction. Secondly, the multi-physics field coupling model takes into account the interaction of electromagnetic fields, thermal fields and flow fields, deepens the understanding of the behavior of the molten pool, and provides a theoretical basis for the optimization of process parameters. Thirdly, the application of the volume fluid method realizes the accurate tracking of the deformation of the free surface of the molten pool, and improves the authenticity of the simulation results. In addition, by systematically changing the electromagnetic field parameters and simulating, a comprehensive molten pool behavior prediction database is established, which lays the foundation for subsequent process optimization. Finally, the closed-loop design of the method enables the prediction results to be directly used to guide the adjustment of process parameters, and realizes the intelligent control of the additive manufacturing process. In general, this method significantly improves the predictability and controllability of the plasma arc additive manufacturing process by integrating data-driven and physical-driven methods, and provides strong technical support for the manufacture of high-quality and high-precision parts.

[0106] Figure 2 FIG. 1 is a schematic diagram of a structure of an electromagnetic field composite powder feeding plasma arc additive manufacturing formability control system according to an embodiment of the present invention. Figure 2 As shown, the system comprises: The first unit is used to obtain the three-dimensional model data of the workpiece to be manufactured and perform layered slicing processing to generate layered slicing data, plan the process path of plasma arc additive manufacturing based on the layered slicing data, control the plasma arc welding gun to move along a preset trajectory according to the process path, and at the same time, convey metal powder to the molten pool area through a powder feeding device, and apply a composite electromagnetic field in the molten pool area, wherein the composite electromagnetic field is generated by a plurality of independently controlled electromagnetic coils, and the current parameters of each electromagnetic coil are dynamically adjusted according to the geometric characteristics of the current manufacturing layer to form a target composite electromagnetic field, and the flow direction and speed of the liquid metal in the molten pool during the deposition process are regulated by the target composite electromagnetic field to control the morphology and internal organizational structure of the deposited layer; The second unit is used to collect process parameters and dynamic image data of the molten pool in the deposition process in real time during the process of regulating the liquid metal in the molten pool, and to analyze the process parameters and the dynamic image data of the molten pool in real time by using a pre-trained state recognition model to identify abnormal deposition states, and to establish a dynamic mapping relationship between process parameters and deposition quality based on the analysis results, and to adaptively adjust the process parameters of subsequent deposition layers according to the dynamic mapping relationship; The third unit is used to measure the actual geometric size and surface morphology of the deposited layer by using a laser scanning system after completing the deposition of each layer, obtain measurement data, compare the measurement data with the theoretical model generated by preset layered slicing data, calculate the geometric error and surface roughness, correct the layered slicing data and process path of the next deposited layer based on the geometric error and surface roughness, analyze the cumulative error trend of the geometric error, optimize the control parameters of the composite electromagnetic field according to the cumulative error trend, execute the deposition manufacturing of the next layer based on the optimized composite electromagnetic field control parameters and the corrected layered slicing data and process path, and execute the cycle until the manufacturing of the entire workpiece is completed.

[0107] 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.

[0108] A fourth aspect of the embodiments of the present invention is: 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.

[0109] 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.

[0110] 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. A method for controlling the formability of electromagnetic field composite powder feeding plasma arc additive manufacturing, characterized in that: include: Acquire the three-dimensional model data of the workpiece to be manufactured and perform layered slicing processing to generate layered slicing data, plan the process path of plasma arc additive manufacturing based on the layered slicing data, control the plasma arc welding gun to move along a preset trajectory according to the process path, and at the same time, convey metal powder to the molten pool area through a powder feeding device, apply a composite electromagnetic field in the molten pool area, the composite electromagnetic field is generated by a plurality of independently controlled electromagnetic coils, and according to the geometric characteristics of the current manufacturing layer, dynamically adjust the current parameters of each electromagnetic coil to form a target composite electromagnetic field, and regulate the flow direction and speed of the liquid metal in the molten pool during the deposition process through the target composite electromagnetic field to control the morphology and internal organizational structure of the deposited layer; In the process of regulating the liquid metal in the molten pool, the process parameters and the dynamic image data of the molten pool in the deposition process are collected in real time, and the process parameters and the dynamic image data of the molten pool are analyzed in real time by using a pre-trained state recognition model to identify abnormal deposition states, and a dynamic mapping relationship between the process parameters and the deposition quality is established based on the analysis results, and the process parameters of the subsequent deposition layers are adaptively adjusted according to the dynamic mapping relationship; After completing the deposition of each layer, a laser scanning system is used to measure the actual geometric dimensions and surface morphology of the deposited layer to obtain measurement data, which are compared with the theoretical model generated by preset layered slicing data to calculate the geometric error and surface roughness. Based on the geometric error and surface roughness, the layered slicing data and process path of the next deposition layer are corrected, the cumulative error trend of the geometric error is analyzed, and according to the cumulative error trend, the control parameters of the composite electromagnetic field are optimized. Based on the optimized composite electromagnetic field control parameters, the corrected layered slicing data and process path, the deposition manufacturing of the next layer is executed, and the cycle is performed until the manufacturing of the entire workpiece is completed.

2. The method according to claim 1, characterized in that Acquiring three-dimensional model data of a workpiece to be manufactured and performing layered slicing processing to generate layered slicing data, planning a process path for plasma arc additive manufacturing based on the layered slicing data, and controlling a plasma arc welding gun to move along a preset trajectory according to the process path includes: Acquire three-dimensional model data of a workpiece to be manufactured, convert the three-dimensional model data into a multi-resolution voxel representation by using an octree algorithm, dynamically adjust the voxel size by calculating a local curvature tensor to obtain a voxelized model, perform adaptive layering processing on the voxelized model, optimize the layer thickness by using a gradient descent method, and generate layered data; Based on the layered data, a phase method topology optimization algorithm is introduced, and the stress distribution of the workpiece is predicted by combining finite element analysis to generate optimized filling structure data. A multi-objective particle swarm optimization algorithm is used to plan the process path of the filling structure data, and a multi-objective function including forming accuracy, surface quality, residual stress and manufacturing efficiency is established to obtain an initial process path. The initial process path is optimized by using a Markov decision process, the path direction between adjacent layers is designed, a three-dimensional balance of residual stress is achieved, and a tree-like support structure is automatically generated for the cantilever structure to obtain an optimized process path; Based on the optimized process path, an inverse kinematics algorithm is used to generate a six-degree-of-freedom motion trajectory instruction of a plasma arc welding gun, and a Bezier curve is introduced to smooth the motion trajectory. In the motion trajectory instruction, a powder feeding amount and arc parameter dynamic adjustment instruction based on fuzzy logic control are embedded to form a comprehensive process instruction. A distributed computing framework is used to convert the comprehensive process instructions into control instructions that can be recognized by the equipment, and deposition process data is collected in real time. The deposition process data is fused using a Kalman filter algorithm to obtain fused real-time status data. According to the control instructions and the fused real-time status data, a model predictive control strategy is adopted to dynamically adjust the motion trajectory and process parameters of the plasma arc welding gun, and the plasma arc welding gun is controlled to move along the motion trajectory through an adaptive control algorithm.

3. The method according to claim 2, characterized in that According to the control instructions and the fused real-time status data, the motion trajectory and process parameters of the plasma arc welding gun are dynamically adjusted by using a model predictive control strategy, and the plasma arc welding gun is controlled to move along the motion trajectory by using an adaptive control algorithm, including: A nonlinear state space model of plasma arc welding process is established, which includes state transfer equations and output equations. The state variables include welding gun position, posture and speed, the control variables include joint driving torque and process parameters, and the output variables include weld geometry characteristics and molten pool state. Based on the nonlinear state space model, a nonlinear model predictive control framework is constructed, and an objective function including an output tracking error and a control increment is defined; A real-time iterative method is used to solve the optimization problem of model predictive control. In each control cycle, model linearization, quadratic programming problem construction and iterative solution are performed to obtain a set of optimized control sequences. Design a model reference adaptive control algorithm, establish a reference model, calculate the adaptive law based on the tracking error between the actual system output and the reference model output, and dynamically adjust the controller parameters; The extended Kalman filter is used to fuse multi-source data from visual sensors, force sensors, and thermocouples, and the prior state estimate is calculated based on the current state estimate and control input. The posterior state estimate is updated using the measured data to obtain the optimal estimate of the system state. The optimal estimate of the system state is input into a nonlinear model predictive controller, an optimized control sequence is calculated, a comprehensive control instruction is generated based on the optimized control sequence and the compensation amount of the adaptive control algorithm, the comprehensive control instruction is applied to the drive system of the plasma arc welding gun, the motion trajectory and process parameters of the welding gun are adjusted, and the plasma arc welding gun is controlled to move according to the adjusted motion trajectory and process parameters.

4. The method according to claim 1, characterized in that: Using a pre-trained state recognition model, the process parameters and the molten pool dynamic image data are analyzed in real time to identify abnormal deposition states including: The multi-sensor fusion system is used to collect the process parameters and molten pool dynamic image data in the deposition process in real time, and the image preprocessing algorithm is used to perform real-time denoising and enhancement on the molten pool dynamic image data to obtain preprocessed molten pool dynamic image data, and the process parameters are normalized to obtain normalized process parameter data; The pre-processed molten pool dynamic image data and the normalized process parameter data are time synchronized and fused to obtain multi-modal input data, and the multi-modal input data is input into a pre-trained multi-label classification network to obtain an abnormal state prediction result; Based on the abnormal state prediction result, a gradient weighted class activation mapping algorithm is used to generate a class activation heat map, the class activation heat map is superimposed with the preprocessed molten pool dynamic image data, and an adaptive threshold segmentation algorithm is used to locate the abnormal area boundary to obtain the abnormal area segmentation result; Calculate the initial confidence score of each abnormal state prediction result, and use random dropout technology to perform multiple forward propagations on the multi-label classification network to calculate the variance of each abnormal state prediction result to obtain the uncertainty score of the model prediction; Based on the initial confidence score and the uncertainty score, a comprehensive confidence score of each abnormal state is calculated, and a time series analysis is performed on the comprehensive confidence scores of the abnormal states of multiple consecutive frames of images to obtain a final confidence score after time series correction; According to the abnormal region segmentation result and the final confidence score after the timing correction, combined with a pre-established abnormal state feature library, the abnormal deposition state, position, area and severity in the plasma arc additive manufacturing process are determined.

5. The method according to claim 4, characterized in that Based on the initial confidence score and the uncertainty score, the formula for calculating the comprehensive confidence score of each abnormal state is as follows: ； Among them, S i，t represents the comprehensive confidence score of abnormal state i at time t, t represents the current time, t0 represents the starting time of the image sequence, β i represents the weight parameter of state i, represents the initial confidence score of abnormal state i, λ represents the uncertainty control coefficient, represents the uncertainty score of abnormal state i, θ represents the uncertainty attenuation coefficient, S i，t-1 represents the comprehensive confidence score of abnormal state i at time t-1, Represents the set of uncertainty scores for all abnormal states.

6. The method according to claim 1, characterized in that The actual geometric dimensions and surface morphology of the deposited layer are measured by a laser scanning system to obtain measurement data, and the measurement data are compared with the theoretical model generated by the preset layered slicing data to calculate the geometric error and surface roughness. Based on the geometric error and surface roughness, the layered slicing data and process path of the next deposited layer are corrected, and the cumulative error trend of the geometric error is analyzed. According to the cumulative error trend, the control parameters of the composite electromagnetic field are optimized, including: A laser scanning system is used to measure the surface of the current layer after deposition to obtain three-dimensional morphology information, the three-dimensional morphology information is compared with a theoretical model generated based on a computer-aided design model and preset layered slice data, and an iterative closest point algorithm is used to perform three-dimensional registration to obtain an actual model after registration; Based on the actual model after registration, the geometric error and surface roughness of the current layer are calculated, where the geometric error includes contour error and thickness error, and the surface roughness is represented by the arithmetic mean roughness value; According to the geometric error, an adaptive algorithm is used to correct the layered slice data of the next deposition layer, the slice profile is adjusted by a reverse offset technology for the profile error, and the thickness error is corrected by dynamically adjusting the interlayer height to obtain the corrected layered slice data; Based on the surface roughness, the process path of the next deposition layer is corrected, and the scanning path and the overlap rate are dynamically adjusted by using an adaptive scanning strategy to obtain a corrected process path; The cumulative trend of the geometric error is analyzed, and a prediction model of the influence of thermal deformation and residual stress on geometric accuracy is established by using a long short-term memory neural network. Based on the prediction model, the influence of the electromagnetic field on the molten pool behavior is simulated by a multi-physical field coupling model, and the molten pool morphology and temperature distribution under different electromagnetic field parameters are predicted to obtain the prediction result of the molten pool behavior; Based on the obtained molten pool behavior prediction results, the static magnetic field intensity, alternating electromagnetic field frequency and intensity are set as optimization variables, and chromosome encoding is constructed. The cumulative geometric error under different parameter combinations is calculated based on the prediction model as the fitness function, and a new generation of population is generated through selection, crossover and mutation operations. The process of evaluating fitness, selecting and generating a new generation of population is repeated until the preset number of iterations is reached, and the individual with the best fitness is selected and decoded to obtain the optimized composite electromagnetic field control parameters.

7. The method according to claim 6, characterized in that A prediction model of the influence of thermal deformation and residual stress on geometric accuracy is established by using a long short-term memory neural network. Based on the prediction model, the influence of the electromagnetic field on the molten pool behavior is simulated by a multi-physics field coupling model, and the molten pool morphology and temperature distribution under different electromagnetic field parameters are predicted. The prediction results of the molten pool behavior include: Based on the cumulative trend of geometric errors, a long short-term memory neural network model is constructed, in which the input layer receives the process parameters, temperature field data and geometric error data of the historical layer, and the output layer predicts the geometric error of the next layer; Using mean square error as the loss function, the constructed long short-term memory neural network model is trained by the optimizer to obtain a trained geometric error prediction model; Constructing a multi-physics field coupling model including electromagnetic field, thermal field and flow field, and setting initial conditions and boundary conditions, obtaining current density and magnetic field distribution by solving electromagnetic field equations in the multi-physics field coupling model, and calculating Joule heat and electromagnetic force; Substitute the obtained Joule heat into the thermal field equation to solve the temperature field distribution, substitute the obtained electromagnetic force into the flow field equation to solve the velocity field and pressure field, update the molten pool morphology and free surface based on the obtained temperature field distribution, velocity field and pressure field, and use the volume fluid method to track the molten pool interface; Repeat the solution of the electromagnetic field, thermal field and flow field equations until the preset time step is reached to obtain the prediction results of the molten pool morphology and temperature distribution. Then change the electromagnetic field parameters and repeat the solution steps of the multi-physics field coupling model to obtain the molten pool morphology and temperature distribution under different electromagnetic field parameters, and finally obtain the prediction results of the molten pool behavior.

8. An electromagnetic field composite powder feeding plasma arc additive manufacturing formability control 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 obtain the three-dimensional model data of the workpiece to be manufactured and perform layered slicing processing to generate layered slicing data, plan the process path of plasma arc additive manufacturing based on the layered slicing data, control the plasma arc welding gun to move along a preset trajectory according to the process path, and at the same time, convey metal powder to the molten pool area through a powder feeding device, and apply a composite electromagnetic field in the molten pool area, wherein the composite electromagnetic field is generated by a plurality of independently controlled electromagnetic coils, and the current parameters of each electromagnetic coil are dynamically adjusted according to the geometric characteristics of the current manufacturing layer to form a target composite electromagnetic field, and the flow direction and speed of the liquid metal in the molten pool during the deposition process are regulated by the target composite electromagnetic field to control the morphology and internal organizational structure of the deposited layer; The second unit is used to collect process parameters and dynamic image data of the molten pool in the deposition process in real time during the process of regulating the liquid metal in the molten pool, and to analyze the process parameters and the dynamic image data of the molten pool in real time by using a pre-trained state recognition model to identify abnormal deposition states, and to establish a dynamic mapping relationship between process parameters and deposition quality based on the analysis results, and to adaptively adjust the process parameters of subsequent deposition layers according to the dynamic mapping relationship; The third unit is used to measure the actual geometric size and surface morphology of the deposited layer by using a laser scanning system after completing the deposition of each layer, obtain measurement data, compare the measurement data with the theoretical model generated by preset layered slicing data, calculate the geometric error and surface roughness, correct the layered slicing data and process path of the next deposited layer based on the geometric error and surface roughness, analyze the cumulative error trend of the geometric error, optimize the control parameters of the composite electromagnetic field according to the cumulative error trend, execute the deposition manufacturing of the next layer based on the optimized composite electromagnetic field control parameters and the corrected layered slicing data and process path, and execute the cycle until the manufacturing of the entire workpiece is completed.

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