Heat input control method and system for arc additive continuous forming

Abstract

The present invention provides a heat input control method and system for arc additive continuous forming, which relates to the field of additive manufacturing technology, including real-time acquisition of process parameters, inputting preprocessing process data into an adaptive heat input calculation model, building it based on thermodynamic principles and heat transfer theory, and calculating the real-time heat input; starting a multi-spectral imaging device, performing real-time temperature field scanning, recording temperature field scanning data, extracting temperature gradient and cooling rate information, and performing correlation calculation with the real-time heat input to determine an impact relationship evaluation result; generating a heat input sequence based on the impact relationship evaluation result; inputting the heat input sequence into an adaptive control system built based on a reinforcement learning algorithm, calculating the optimal parameter scheme based on preset target deposition layer performance indicators and the current temperature field state, adjusting the corresponding time and space distribution of the control energy input based on the optimal parameter scheme, and controlling the heat input of different positions and different layers of the deposition layer.

Description

Technical Field

[0001] The present invention relates to the field of additive manufacturing technology, and in particular to a heat input control method and system for arc additive continuous forming. Background Art

[0002] Arc additive manufacturing is a process that uses an electric arc as a heat source to melt metal wires, and then solidifies and accumulates them on a computer-controlled three-dimensional motion platform according to a planned path to form the required parts. It has the advantages of low cost, high efficiency, and the ability to manufacture large parts. It has broad application prospects in the fields of aerospace, automobile manufacturing, etc. However, heat input control in the arc additive manufacturing process has always been one of the main challenges facing this technology, which directly affects the quality, geometric accuracy and performance of the formed parts.

[0003] Existing arc additive manufacturing heat input control methods often use empirical models or simplified physical models to predict and control heat input. There are some limitations. Arc additive manufacturing involves complex heat conduction, phase change and mechanical behavior, and it is difficult to accurately capture the complex thermal-mechanical coupling process; traditional methods usually require a lot of experiments and iterations to find the right combination of process parameters, and parameter optimization is inefficient; due to the dynamic changes in the manufacturing environment and the inhomogeneity of materials, there is a lack of real-time adaptive capabilities.

[0004] With the development of artificial intelligence technology, methods based on deep reinforcement learning have made significant progress in the field of complex system control. However, the application of deep reinforcement learning to the heat input control of arc additive manufacturing still faces many challenges. Therefore, there is an urgent need for a heat input control method for arc additive continuous forming based on deep reinforcement learning, which can make full use of historical data, achieve multi-objective optimization, and have real-time adaptive capabilities to improve the forming quality and efficiency of arc additive manufacturing. The present invention can solve the problems in the prior art. Summary of the invention

[0005] The embodiments of the present invention provide a method and system for controlling heat input in arc additive continuous forming, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] A method for controlling heat input in arc additive continuous forming is provided, comprising:

[0008] The process parameters in the arc additive continuous forming process are collected in real time by multiple sensors, and pre-processing process data are obtained by filtering and standardizing the process parameters; the pre-processing process data is input into an adaptive heat input calculation model, and the adaptive heat input calculation model is constructed based on thermodynamic principles and heat transfer theory to calculate the real-time heat input in the deposition process;

[0009] Based on the real-time heat input, a multispectral imaging device is started to perform a real-time temperature field scan on the surface of the deposited layer, and the temperature field scan data is recorded. The temperature gradient and cooling rate information of each region of the deposited layer are extracted through image processing algorithm analysis, and the temperature gradient and cooling rate information are correlated with the real-time heat input to determine the impact relationship evaluation result of the current heat input state on the microstructure and mechanical properties of the deposited layer; and a heat input sequence is generated according to the impact relationship evaluation result;

[0010] The heat input sequence is input into an adaptive control system built based on a reinforcement learning algorithm, and the optimal parameter scheme is calculated according to the preset target deposition layer performance indicators and the current temperature field state. Based on the optimal parameter scheme, the process parameters are adjusted in real time through a digital welding power supply and a high-precision wire feeding system, and the arc pulse modulation device is started at the same time. The corresponding spatiotemporal distribution of the controlled energy input is adjusted according to the optimal parameter scheme, and the heat input to different positions and layers of the deposition layer is controlled based on the spatiotemporal distribution.

[0011] In an optional embodiment,

[0012] The pretreatment process data is input into an adaptive heat input calculation model. The adaptive heat input calculation model is constructed based on thermodynamic principles and heat transfer theory. The real-time heat input during the deposition process is calculated to include:

[0013] Construct multi-scale nodes, define microstructure features as microscale nodes, cross-sectional grid points as mesoscale nodes, and the centers of surface partitioned areas as macroscale nodes;

[0014] Based on the heat conduction theory, the spatial connection between multi-scale nodes is established through physical distance; based on different time steps, the temporal connection between multi-scale nodes is established; based on each scale in the multi-scale nodes, cross-scale connections are established to form a multi-scale space-time graph;

[0015] Performing a crystal graph convolution operation on the micro-scale nodes, performing an anisotropic graph convolution operation on the meso-scale nodes, and performing a spherical graph convolution operation on the macro-scale nodes to obtain multi-scale features; bidirectionally transmitting the multi-scale features through upsampling and downsampling operations, adjusting the scale importance through an attention mechanism, and performing scale fusion to obtain fused multi-scale features;

[0016] Processing the fused multi-scale features through a graph converter to obtain long-term temporal dependency information, inputting the long-term temporal dependency information into a graph differential equation, simulating a dynamic process of heat conduction, and obtaining a preliminary heat input prediction result;

[0017] Based on the preliminary heat input prediction result, calculate the material phase change latent heat compensation, convert the material constitutive equation into a Neural Ordinary Differential Equation, use the differentiable yield criterion to calculate the material property calculation result, combine the above material property calculation result with the preliminary heat input prediction result, and obtain a revised heat input prediction result;

[0018] Based on the corrected heat input prediction results, a loss function including energy conservation constraints, thermodynamic law constraints and difficult sample loss terms is constructed, a gradient penalty term is added, and the weights of the corresponding terms in the loss function are dynamically adjusted based on the different stages of the arc additive forming process.

[0019] Based on the goal of minimizing the loss function, the adaptive heat input calculation model is iteratively trained to output the final real-time heat input prediction result.

[0020] In an optional embodiment,

[0021] Through image processing algorithm analysis, the temperature gradient and cooling rate information of each area of ​​the deposit layer are extracted, and the temperature gradient and cooling rate information are correlated with the real-time heat input to determine the impact of the current heat input state on the microstructure and mechanical properties of the deposit layer. The evaluation results include:

[0022] Performing image preprocessing on the temperature field scanning data, including using an adaptive median filtering algorithm to remove thermal noise, and using a contrast-limited adaptive histogram equalization method to enhance image contrast, to obtain preprocessed temperature field data;

[0023] The temperature gradient of the preprocessed temperature field data is calculated using a multi-scale gradient, and the gradient is calculated at different spatial scales and weighted fusion is performed to obtain a total temperature gradient field; the cooling rate of each pixel is calculated by analyzing the temperature change between consecutive frames in the preprocessed temperature field data, and the cooling rate is smoothed using a Kalman filter to obtain smoothed cooling rate data;

[0024] Using a long short-term memory network to establish a dynamic mapping relationship between the real-time heat input and the total temperature gradient field and the smoothed cooling rate data, a heat input-temperature field characteristic mapping model is obtained;

[0025] The heat input-temperature field characteristic mapping model is used, based on the total temperature gradient field and the smoothed cooling rate data, and a multiphase field model of coupled temperature field and concentration field is used to predict the grain size, grain morphology and grain orientation distribution of the deposited layer; the mechanical properties of the deposited layer, including yield strength, tensile strength, elongation and hardness, are calculated using a multiscale modeling method based on machine learning, in combination with the grain size, grain morphology and grain orientation distribution, using a pre-established microstructure-performance relationship database;

[0026] Based on the mechanical properties of the deposited layer and the grain size, grain morphology and grain orientation distribution, an evaluation result of the influence of the heat input state on the microstructure and mechanical properties of the deposited layer is generated.

[0027] In an optional embodiment,

[0028] Based on the impact relationship evaluation results, the heat input sequence is generated including:

[0029] Based on the results of the impact relationship assessment, the heat input is used as the decision variable, discretized into a time series, and the microstructure characteristics and mechanical properties are used as optimization targets to construct a multi-objective optimization model.

[0030] Using the Latin hypercube sampling method to generate an initial population in the decision space of the multi-objective optimization model, and using a deep neural network to construct a mapping relationship from a hot input sequence to the optimization objective to form a proxy model;

[0031] According to a preset population division ratio, the initial population is randomly divided into a first subgroup and a second subgroup, and computing resources are allocated accordingly, including: using a heat input-temperature field feature mapping model and a multiphase field model to perform accuracy evaluation on individuals in the first subgroup, and using the proxy model to perform rapid evaluation on individuals in the second subgroup; determining an objective function value for each individual in the initial population, and based on the obtained objective function value, using a fast non-dominated sorting algorithm to stratify the initial population, obtaining stratification results, and calculating the crowding distance of each individual;

[0032] According to the stratification result and crowding distance of the initial population, in combination with the current iteration stage, the parameters of the evolutionary operator are dynamically adjusted, and the alternative probabilities corresponding to the preset multiple evolutionary operators are updated by an adaptive strategy, and the evolutionary operator is determined by a roulette selection algorithm; based on the tournament selection method, the parent individual is selected from the initial population using the stratification result and crowding distance, and the determined evolutionary operator is applied to generate the offspring individual, the offspring individual is evaluated to obtain the objective function value corresponding to each offspring individual, the parent individual and the offspring individual are merged to form a merged population, non-dominated sorting is performed based on the objective function value of each individual in the merged population, and the next generation population is selected from the merged population in combination with the crowding distance; based on the next generation population, in the target space, a point inferior to all non-dominated solutions is selected as a reference point, the exclusive hypervolume of each individual is calculated, the hypervolume contribution of each individual is determined, and the computing resources are allocated according to the hypervolume contribution;

[0033] The iteration is repeated until the preset total number of iterations is reached to obtain the final non-dominated solution set, and the TOPSIS method is applied to select the best compromise solution based on the objective function value of each solution in the final non-dominated solution set to obtain the optimal heat input sequence.

[0034] In an optional embodiment,

[0035] The heat input sequence is input into the adaptive control system built based on the reinforcement learning algorithm. According to the preset target deposition layer performance index and the current temperature field state, the optimal parameter scheme is calculated, including:

[0036] The heat input sequence is discretized and normalized at fixed time intervals, and a sliding window is used to treat multiple consecutive time steps as an input unit;

[0037] Use a three-dimensional convolutional neural network to compress the temperature field information, convert the target performance index into a vector representation, encode the current layer geometric characteristics and material thermophysical parameters, and construct the state space;

[0038] Based on the state space, a hybrid action space is determined, including a discrete action space and a continuous action space, wherein the discrete action space includes current, voltage, wire feeding speed, pulse frequency and duty cycle, and the continuous action space is used to fine-grainedly adjust corresponding parameters in the discrete action space;

[0039] Construct a multi-objective reward function, including quality reward, geometric accuracy reward, energy efficiency reward and stability reward, and use weight coefficients to balance the reward function, where the quality reward is calculated based on the deviation of hardness and strength from the target value, the geometric accuracy reward is determined based on the evaluation of shape accuracy and surface roughness, the energy efficiency reward evaluates the heat input utilization and energy loss, and the stability reward is based on arc stability and spatter degree;

[0040] Initialize the main Q network and the target Q network, wherein the main Q network and the target Q network have the same network structure and different initial parameter values; design a dual-path deep Q network architecture, and divide the output layers of the main Q network and the target Q network into a state value function branch and an advantage function branch, wherein the state value function branch estimates the state value, and the advantage function branch estimates the action advantage; construct an experience playback buffer to store experience samples of the interaction between the agent and the environment, wherein each experience sample contains the current state, the executed action, the reward obtained, the next state, and the termination flag;

[0041] Implement a multi-step bootstrapping mechanism to calculate the target Q value by accumulating discounted rewards over multiple steps and using the target Q network to estimate the state value after multiple steps;

[0042] defining a loss function that combines a temporal difference error and a dual-path architecture to calculate a difference between a Q value predicted by the main Q network and a target Q value;

[0043] Using the samples in the experience replay buffer, calculating the value of the loss function, and updating the parameters of the main Q network based on the value of the loss function;

[0044] Copying the parameters of the master Q network to the target Q network according to a predefined update frequency;

[0045] Repeat until the number of training times reaches the preset value.

[0046] Using the soft actor-critic algorithm, based on the trained reinforcement learning model, the current state is input, including the processed thermal input sequence, temperature field information and target performance indicators, the trained Q network is used to evaluate the value of different actions, and the optimal action is selected through soft maximization operation;

[0047] Through multiple iterations of optimization, the optimal parameter solution is calculated.

[0048] In an optional embodiment,

[0049] Implementing a multi-step bootstrapping mechanism to calculate the target Q value by accumulating discounted rewards over multiple steps and using the target Q network to estimate the state value after multiple steps includes:

[0050] Set a step number parameter to determine the number of time steps to consider when calculating the target Q value; collect and store state, action, reward, and next state information at each time step until continuous experience that meets the step number parameter is collected;

[0051] For the current time step, calculate a multi-step reward value, which includes all rewards from the current time step to the future step number parameter minus one time step, and the estimated Q value of the last time step, where the corresponding reward of each time step is multiplied by the corresponding discount factor;

[0052] Based on the multi-step reward values, a target Q value is determined; a difference between the target Q value and the current Q value is calculated to determine a time difference error, and based on the time difference error, a gradient descent method is used to minimize the mean square error loss and update the main Q network parameters;

[0053] The sliding window method is used to continuously update multi-step experience. At the beginning of a new time step, the earliest experience in the experience replay buffer is removed and the new experience is added to the experience replay buffer.

[0054] According to a second aspect of the embodiments of the present invention,

[0055] A heat input control system for arc additive continuous forming is provided, comprising:

[0056] The first unit is used to collect process parameters in the arc additive continuous forming process in real time through multiple sensors, obtain pre-processing process data by filtering and standardizing the process parameters; input the pre-processing process data into an adaptive heat input calculation model, the adaptive heat input calculation model is constructed based on thermodynamic principles and heat transfer theory, and calculates the real-time heat input in the deposition process;

[0057] The second unit is used to start the multispectral imaging device based on the real-time heat input, perform real-time temperature field scanning on the surface of the sediment layer, record the temperature field scanning data, extract the temperature gradient and cooling rate information of each region of the sediment layer through image processing algorithm analysis, associate the temperature gradient and cooling rate information with the real-time heat input, and determine the impact relationship evaluation result of the current heat input state on the microstructure and mechanical properties of the sediment layer; generate a heat input sequence according to the impact relationship evaluation result;

[0058] The third unit is used to input the heat input sequence into the adaptive control system built based on the reinforcement learning algorithm, and calculate the optimal parameter scheme according to the preset target deposition layer performance indicators and the current temperature field state; based on the optimal parameter scheme, the process parameters are adjusted in real time through the digital welding power supply and the high-precision wire feeding system, and the arc pulse modulation device is started at the same time. The corresponding time and space distribution of the control energy input is adjusted according to the optimal parameter scheme, and the heat input of different positions and different layers of the deposition layer is controlled based on the time and space distribution.

[0059] According to a third aspect of the embodiments of the present invention,

[0060] An electronic device is provided, comprising:

[0061] processor;

[0062] a memory for storing processor-executable instructions;

[0063] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0064] A fourth aspect of the embodiments of the present invention is:

[0065] 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.

[0066] In an embodiment of the present invention, an adaptive heat input calculation model uses thermodynamic principles and heat transfer theory to accurately calculate the actual heat input during the deposition process, ensuring that the material is heated evenly and avoiding problems such as overheating or insufficient heat, thereby improving the forming quality of the deposited layer; by combining real-time collected data and an adaptive model, the heat input can be dynamically adjusted according to different process conditions, thereby improving the flexibility and adaptability of the entire manufacturing process and ensuring the stability of process quality under different deposition conditions; the temperature field of the deposited layer is scanned in real time by a multi-spectral imaging device to obtain accurate temperature gradient and cooling rate data, thereby ensuring detailed monitoring of the thermal distribution of the material during the manufacturing process and reducing defects caused by temperature runaway; the temperature field data is quickly analyzed using an image processing algorithm to timely extract key thermal parameters; based on the temperature field analysis results, a new heat input sequence is dynamically adjusted and generated so that subsequent heat input more accurately matches the actual needs of the material, thereby ensuring the consistency of the quality of the deposited layer and the optimization of mechanical properties; an adaptive control system based on a reinforcement learning algorithm can calculate the optimal process parameters through continuous learning and adjustment. The system can automatically optimize operations based on target performance and real-time temperature field status, thereby improving the intelligence level of the process, reducing human intervention, and enhancing the efficiency and accuracy of the manufacturing process. The adaptive control system driven by reinforcement learning forms a closed-loop control mechanism through real-time data feedback and adjustment, ensuring the stability of heat input throughout the entire manufacturing process, reducing the impact of external environment or equipment fluctuations, and further improving the stability and reliability of the process. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 A schematic flow chart of a method for controlling heat input in arc additive continuous forming according to an embodiment of the present invention;

[0068] Figure 2 Schematic diagram of the structure of the heat input control system for arc additive continuous forming according to an embodiment of the present invention. DETAILED DESCRIPTION

[0069] 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.

[0070] 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.

[0071] Figure 1 FIG. 1 is a flow chart of a heat input control method for arc additive continuous forming according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0072] S101. Real-time acquisition of process parameters in the arc additive continuous forming process by multiple sensors, filtering and standardizing the process parameters to obtain pre-processing process data; inputting the pre-processing process data into an adaptive heat input calculation model, wherein the adaptive heat input calculation model is constructed based on thermodynamic principles and heat transfer theory to calculate the real-time heat input in the deposition process;

[0073] In this embodiment, multiple sensors are used to obtain process parameters (such as current, voltage, wire feeding speed, etc.) in real time to ensure that each link in the manufacturing process can be monitored in a timely manner, thereby improving the response speed of the system and being able to adjust the process parameters more quickly according to the real-time situation; filtering is used to eliminate noise interference, and process data is standardized to ensure the stability and consistency of input data, thereby effectively reducing the impact of measurement errors and data fluctuations on process control, thereby improving the accuracy of heat input calculation; the adaptive heat input calculation model uses thermodynamic principles and heat transfer theory to accurately calculate the actual heat input during the deposition process, ensuring that the material is heated evenly, avoiding problems such as overheating or insufficient heat, thereby improving the forming quality of the deposited layer; by combining real-time collected data and adaptive models, the heat input can be dynamically adjusted according to different process conditions, thereby improving the flexibility and adaptability of the entire manufacturing process and ensuring the stability of process quality under different deposition conditions.

[0074] S102. Based on the real-time heat input, start the multispectral imaging device to perform real-time temperature field scanning on the surface of the deposited layer, record the temperature field scanning data, extract the temperature gradient and cooling rate information of each region of the deposited layer through image processing algorithm analysis, associate the temperature gradient and cooling rate information with the real-time heat input, and determine the impact relationship evaluation result of the current heat input state on the microstructure and mechanical properties of the deposited layer; generate a heat input sequence according to the impact relationship evaluation result;

[0075] In this embodiment, the temperature field of the deposited layer is scanned in real time by a multispectral imaging device to obtain accurate temperature gradient and cooling rate data, thereby ensuring detailed monitoring of the thermal distribution of the material during the manufacturing process and reducing defects caused by temperature runaway; the temperature field data is quickly analyzed by an image processing algorithm, so that key thermal parameters (temperature gradient, cooling rate) can be extracted in time, providing an accurate basis for subsequent heat input regulation, thereby improving the efficiency and accuracy of data processing; the temperature field information is associated with the real-time heat input, so that the influence of heat input on the microstructure and mechanical properties of the material can be evaluated, and heat input problems that may affect the material performance can be discovered and corrected in time during the manufacturing process; based on the temperature field analysis results, a new heat input sequence is dynamically adjusted and generated, so that subsequent heat input can more accurately match the actual needs of the material, thereby ensuring the consistency of the quality of the deposited layer and the optimization of the mechanical properties; by analyzing the correlation between the temperature gradient, cooling rate and heat input, the microstructure and mechanical properties of the material can be better controlled, and the reliability of the manufacturing process and product quality can be improved.

[0076] S103. Input the heat input sequence into the adaptive control system constructed based on the reinforcement learning algorithm, and calculate the optimal parameter scheme according to the preset target deposition layer performance indicators and the current temperature field state; based on the optimal parameter scheme, adjust the process parameters in real time through the digital welding power supply and high-precision wire feeding system, and start the arc pulse modulation device at the same time, adjust the corresponding time and space distribution of the control energy input according to the optimal parameter scheme, and control the heat input to different positions and layers of the deposition layer based on the time and space distribution.

[0077] In this embodiment, the adaptive control system based on the reinforcement learning algorithm can calculate the optimal process parameters through continuous learning and adjustment. The system can automatically optimize the operation according to the target performance and the real-time temperature field state, improve the intelligent level of the process, reduce manual intervention, and improve the efficiency and accuracy of the manufacturing process; through the digital welding power supply and high-precision wire feeding system, the process parameters can be adjusted in real time and meticulously. Precisely control key parameters such as current, voltage, and wire feeding speed to ensure that the forming process of each layer of material is maintained in the optimal state, thereby improving the forming quality and consistency of the deposited layer. The system can adjust the spatiotemporal distribution of energy input through the arc pulse modulation device according to the thermal requirements of different levels and positions, ensuring that the heat input in different areas is precisely controlled to avoid problems such as material deformation and stress concentration caused by uneven heat. By dynamically adjusting the spatiotemporal distribution of heat input, the heat input of the deposited layer at different positions and levels is ensured to be optimal, which can effectively improve the geometric accuracy, microstructural uniformity, and mechanical property consistency of the deposited layer, thereby improving the overall product quality and reliability. The adaptive control system driven by reinforcement learning forms a closed-loop control mechanism through real-time data feedback and adjustment, ensuring the stability of heat input throughout the manufacturing process, reducing the impact of external environment or equipment fluctuations, and further improving the stability and reliability of the process.

[0078] An optional implementation is as follows:

[0079] First, multiple sensors are used to collect the process parameters of the arc additive continuous forming process in real time. These sensors include voltage sensors, current sensors, wire feed speed sensors, welding speed sensors, etc. The voltage sensor collects real-time arc voltage, the current sensor collects real-time arc current, the wire feed speed sensor collects real-time wire feed speed, and the welding speed sensor collects real-time welding speed. These sensors continuously collect data at a sampling frequency of 100Hz.

[0080] The collected raw process parameter data is filtered and standardized to obtain pre-processed process data. Specifically, a bandpass filter is used to filter the raw data to remove high-frequency noise and low-frequency drift. The filtered data is then normalized to unify the parameters of different dimensions into the range of 0-1. For example, voltage data is normalized within the range of 0-40V, current data is normalized within the range of 0-300A, wire feeding speed is normalized within the range of 0-15m / min, and welding speed is normalized within the range of 0-10mm / s.

[0081] The pretreatment process data is input into the adaptive heat input calculation model. The model is constructed based on thermodynamic principles and heat transfer theory, taking into account multiple heat sources such as arc heat, wire melting heat, droplet overheating heat, as well as multiple heat losses such as heat conduction, convection and radiation. The model inputs process parameters such as voltage, current, wire feeding speed and welding speed, and the output is real-time heat input. For example, at a certain moment, the input voltage is 25V, the current is 180A, the wire feeding speed is 8m / min, and the welding speed is 5mm / s. The calculated real-time heat input is 2100J / s.

[0082] Based on the calculated real-time heat input, the multispectral imaging device is started to scan the temperature field of the sediment surface in real time. The multispectral imaging device includes a visible light camera and an infrared thermal imager, which can simultaneously collect visible light images and thermal images. The resolution of the visible light camera is 1920×1080 pixels and the frame rate is 60fps. The resolution of the infrared thermal imager is 640×480 pixels, the temperature measurement range is 0-1500℃, the temperature measurement accuracy is ±2℃, and the frame rate is 30fps.

[0083] The collected temperature field scanning data is processed and analyzed to extract the temperature gradient and cooling rate information of each area of ​​the sediment layer. First, the thermal image is denoised and enhanced, and then the edge detection algorithm is used to extract the sediment layer contour. The sediment layer area is divided into multiple sub-areas, and the average temperature of each sub-area is calculated. The temperature gradient is calculated by the temperature difference and distance between adjacent sub-areas, and the cooling rate is calculated by the temperature change between consecutive frames. For example, at a certain moment, the temperature in the center area of ​​the sediment layer is 1200℃, the temperature in the edge area is 800℃, the temperature gradient is 80℃ / mm, and the cooling rate is 50℃ / s.

[0084] The extracted temperature gradient and cooling rate information are correlated with the real-time heat input to determine the influence of the current heat input state on the microstructure and mechanical properties of the deposited layer. A mapping relationship between heat input, temperature gradient, cooling rate, microstructure and mechanical properties is established. For example, when the heat input is 2100 J / s, the temperature gradient is 80°C / mm, and the cooling rate is 50°C / s, the predicted grain size of the deposited layer is 50 μm and the yield strength is 350 MPa.

[0085] According to the impact relationship assessment results, a heat input sequence is generated. The heat input sequence contains the target heat input required for different positions and layers. For example, for the bottom layer, the target heat input in the center area is 2200 J / s, and the target heat input in the edge area is 1800 J / s; for the top layer, the target heat input in the center area is 2000 J / s, and the target heat input in the edge area is 1600 J / s.

[0086] The generated heat input sequence is input into an adaptive control system built based on a reinforcement learning algorithm. The system uses a deep Q network (DQN) algorithm, takes the current temperature field state as input, and outputs the optimal process parameter adjustment plan. The system's reward function is defined based on the difference between the target deposited layer performance index and the current performance. For example, the target grain size is 40μm and the target yield strength is 400MPa.

[0087] Based on the calculated optimal parameter solution, the process parameters are adjusted in real time through the digital welding power supply and high-precision wire feeding system. The digital welding power supply can respond to the current and voltage adjustment instructions within 1ms, with an adjustment accuracy of ±1A and ±0.1V. The speed regulation response time of the high-precision wire feeding system is less than 10ms, and the speed regulation accuracy is ±0.1m / min. For example, the optimal parameter solution at a certain moment is: the current is adjusted to 190A, the voltage is adjusted to 26V, and the wire feeding speed is adjusted to 8.5m / min.

[0088] At the same time, the arc pulse modulation device is started to adjust the corresponding time and space distribution of the control energy input according to the optimal parameter scheme. The arc pulse modulation device can realize millisecond-level current and voltage pulse modulation, with a pulse frequency range of 1-1000Hz and a duty cycle adjustable in the range of 10%-90%. For example, for the edge area, a pulse modulation scheme of 200Hz and 30% duty cycle is used to reduce heat input.

[0089] Based on the calculated spatiotemporal distribution, differentiated heat input control is implemented for different positions and layers of the deposition layer. For example, a continuous constant flow method is used for the center area of ​​the bottom layer, with a heat input of 2200 J / s; a pulse method is used for the edge area of ​​the bottom layer, with an average heat input of 1800 J / s; a low-frequency pulse method is used for the center area of ​​the top layer, with an average heat input of 2000 J / s; and a high-frequency pulse method is used for the edge area of ​​the top layer, with an average heat input of 1600 J / s.

[0090] Through the closed-loop control of the above steps, precise regulation of heat input in the arc additive continuous forming process is achieved, and the thermal history parameters such as the temperature field distribution and cooling rate of the deposited layer can be effectively controlled, thereby achieving directional regulation of the microstructure and mechanical properties of the deposited layer and improving the overall performance and quality stability of the formed parts.

[0091] In an optional embodiment, the pretreatment process data is input into an adaptive heat input calculation model, which is constructed based on thermodynamic principles and heat transfer theory, and the real-time heat input in the deposition process is calculated to include:

[0092] Construct multi-scale nodes, define microstructure features as microscale nodes, cross-sectional grid points as mesoscale nodes, and the centers of surface partitioned areas as macroscale nodes;

[0093] Based on the heat conduction theory, the spatial connection between multi-scale nodes is established through physical distance; based on different time steps, the temporal connection between multi-scale nodes is established; based on each scale in the multi-scale nodes, cross-scale connections are established to form a multi-scale space-time graph;

[0094] Performing a crystal graph convolution operation on the micro-scale nodes, performing an anisotropic graph convolution operation on the meso-scale nodes, and performing a spherical graph convolution operation on the macro-scale nodes to obtain multi-scale features; bidirectionally transmitting the multi-scale features through upsampling and downsampling operations, adjusting the scale importance through an attention mechanism, and performing scale fusion to obtain fused multi-scale features;

[0095] Processing the fused multi-scale features through a graph converter to obtain long-term temporal dependency information, inputting the long-term temporal dependency information into a graph differential equation, simulating a dynamic process of heat conduction, and obtaining a preliminary heat input prediction result;

[0096] Based on the preliminary heat input prediction result, calculate the material phase change latent heat compensation, convert the material constitutive equation into a Neural Ordinary Differential Equation, use the differentiable yield criterion to calculate the material property calculation result, combine the above material property calculation result with the preliminary heat input prediction result, and obtain a revised heat input prediction result;

[0097] Based on the corrected heat input prediction results, a loss function including energy conservation constraints, thermodynamic law constraints and difficult sample loss terms is constructed, a gradient penalty term is added, and the weights of the corresponding terms in the loss function are dynamically adjusted based on the different stages of the arc additive forming process.

[0098] Based on the goal of minimizing the loss function, the adaptive heat input calculation model is iteratively trained to output the final real-time heat input prediction result.

[0099] The crystal graph convolution operation specifically refers to an operation method based on the microscopic crystal structure of the material, which is mainly used to process micro-scale nodes. It combines the microscopic structural characteristics of the material (such as crystal orientation, grain boundaries, etc.) with the information of neighboring nodes through graph convolution to capture the complex interactions and heat conduction behaviors within the crystal structure. This operation can help accurately describe the heat conduction path inside the material and the corresponding microstructural changes.

[0100] The anisotropic graph convolution operation is specifically used for mesoscale nodes, which can take into account the differences in physical properties in different directions (anisotropy), such as the changes in the thermal conductivity of materials in different directions. Through this convolution operation, the uneven distribution of heat in different directions on the mesoscale can be more accurately captured, which is especially suitable for the complex transfer of heat conduction on cross-sectional grid points.

[0101] The spherical graph convolution operation is specifically used for macro-scale nodes to process large-scale regional features. This operation is based on the convolution operation of the spherical coordinate system, which can effectively capture the characteristics of the center point of the macro-surface divided area and the overall heat transfer of the area. It helps to describe how heat flows between different areas on a large scale, thereby reflecting the heat conduction behavior on a macro scale.

[0102] The material phase change latent heat compensation specifically refers to the fact that the material absorbs or releases a large amount of latent heat during the phase change process, which has a great impact on the heat input. The calculation of the material phase change latent heat compensation takes this into account, and compensates for the latent heat effect in the heat input model by accurately calculating the phase change behavior of the material. This ensures that the heat input prediction results are more accurate when the material changes phase, thereby better simulating the actual heat conduction process.

[0103] First, a multi-scale node network is constructed. Microstructure features such as grain size and phase composition are defined as microscale nodes; cross-sectional grid points are defined as mesoscale nodes; and the center points of surface partition areas are defined as macroscale nodes. For example, a 10μm-scale grain can be defined as a microscale node, a 1mm grid can be defined as a mesoscale node, and the center of a 10mm area can be defined as a macroscale node.

[0104] Then, a multi-scale space-time graph is established. Based on the heat conduction theory, the spatial connection between multi-scale nodes is established through physical distance, such as the connection between adjacent grains, adjacent grid points, and adjacent regions. Based on different time steps, the time connection between nodes is established, such as the connection between the nodes at the same position at time t and time t+Δt. Based on each scale in the multi-scale node, a cross-scale connection is established, such as the connection between a microscopic node and its mesoscopic node. Finally, a multi-scale space-time graph containing three types of connections: space, time, and cross-scale is formed.

[0105] Next, the corresponding graph convolution operations are performed on nodes of different scales. Crystal graph convolution operations are performed on micro-scale nodes to consider the anisotropy of the crystal structure; anisotropic graph convolution operations are performed on meso-scale nodes to consider the anisotropic thermal conductivity of the material; and spherical graph convolution operations are performed on macro-scale nodes to consider the effect of surface curvature on thermal conduction. These operations can be used to obtain feature representations of each scale.

[0106] Then, multi-scale feature fusion is performed. The microscopic features are transferred to the mesoscopic and macroscopic scales through upsampling operations, and the macroscopic features are transferred to the mesoscopic and microscopic scales through downsampling operations, thus realizing bidirectional feature transfer. At the same time, the attention mechanism is introduced to dynamically adjust the importance weights of each scale according to the current process parameters. Finally, the features of each scale are weighted and fused to obtain the fused multi-scale feature representation.

[0107] Next, the graph transformer is used to process the fused multi-scale features to capture long-term temporal dependency information. The long-term temporal dependency information is input into the graph differential equation to simulate the dynamic process of heat conduction and obtain preliminary prediction results of heat input. For example, the change of temperature at each location over time during the movement of the heat source can be predicted.

[0108] Based on the preliminary prediction results, the influence of the latent heat of material phase change is further considered. The latent heat absorbed or released by the material during the phase change process is calculated as a compensation term for the heat input. At the same time, the material constitutive equation is converted into the form of a magical ordinary differential equation, and the change in the mechanical properties of the material at high temperature is calculated using a differentiable yield criterion. The material property calculation results are combined with the preliminary heat input prediction results to obtain the revised heat input prediction results.

[0109] Finally, a loss function is constructed and the model is iteratively trained. The loss function contains energy conservation constraints, thermodynamic law constraints, and difficult sample loss terms. The energy conservation constraint term ensures that the prediction result satisfies the energy balance; the thermodynamic law constraint term ensures that the prediction result conforms to the second law of thermodynamics; the difficult sample loss term increases the attention to difficult-to-predict samples. At the same time, a gradient penalty term is added to improve the generalization ability of the model. Based on the different stages of the arc additive forming process, such as preheating, stable deposition, cooling, etc., the weights of the corresponding terms in the loss function are dynamically adjusted. By minimizing the loss function, the adaptive heat input calculation model is iteratively trained, and finally the real-time heat input prediction results are output.

[0110] In practical applications, this method can accurately predict the real-time heat input during arc additive manufacturing. For example, for arc additive manufacturing of Ti-6Al-4V alloy, under the process parameters of 60A current, 12V voltage, and 6mm / s wire feeding speed, this method can predict the temperature field distribution at each position during the movement of the heat source, as well as the phase change and mechanical property changes of the material. With this information, the process parameters can be optimized to improve the forming quality.

[0111] In this embodiment, by constructing micro-, meso-, and macro-scale nodes and combining them with heat conduction theory, the heat transfer behavior at different scales is accurately described, thereby improving the accuracy of heat input calculation; different types of convolution operations (crystal graph convolution, anisotropic graph convolution, spherical graph convolution) are used to optimize the characteristics of different scales, thereby improving the effectiveness and accuracy of feature extraction; through upsampling, downsampling, and attention mechanisms, bidirectional transmission and fusion of multi-scale features are achieved, thereby enhancing the overall data utilization and prediction accuracy; a graph transformer is used to capture long-term temporal dependencies, thereby improving the modeling capability of the dynamic process of heat conduction; the latent heat of material phase change is considered, combined with the Neural Ordinary Differential Equation, the thermal behavior of the material is accurately simulated, thereby improving the physical consistency of the prediction results; energy conservation and thermodynamic constraints are introduced, and the weights are dynamically adjusted to ensure the stability and accuracy of the model optimization process; through iterative training, the real-time heat input prediction is continuously optimized, thereby improving the accuracy and stability in the additive manufacturing process.

[0112] In an optional embodiment, the temperature gradient and cooling rate information of each region of the deposited layer are extracted through image processing algorithm analysis, and the temperature gradient and cooling rate information are correlated with the real-time heat input to determine the influence of the current heat input state on the microstructure and mechanical properties of the deposited layer. The evaluation results include:

[0113] Performing image preprocessing on the temperature field scanning data, including using an adaptive median filtering algorithm to remove thermal noise, and using a contrast-limited adaptive histogram equalization method to enhance image contrast, to obtain preprocessed temperature field data;

[0114] The temperature gradient of the preprocessed temperature field data is calculated using a multi-scale gradient, and the gradient is calculated at different spatial scales and weighted fusion is performed to obtain a total temperature gradient field; the cooling rate of each pixel is calculated by analyzing the temperature change between consecutive frames in the preprocessed temperature field data, and the cooling rate is smoothed using a Kalman filter to obtain smoothed cooling rate data;

[0115] Using a long short-term memory network to establish a dynamic mapping relationship between the real-time heat input and the total temperature gradient field and the smoothed cooling rate data, a heat input-temperature field characteristic mapping model is obtained;

[0116] The heat input-temperature field characteristic mapping model is used, based on the total temperature gradient field and the smoothed cooling rate data, and a multiphase field model of coupled temperature field and concentration field is used to predict the grain size, grain morphology and grain orientation distribution of the deposited layer; the mechanical properties of the deposited layer, including yield strength, tensile strength, elongation and hardness, are calculated using a multiscale modeling method based on machine learning, in combination with the grain size, grain morphology and grain orientation distribution, using a pre-established microstructure-performance relationship database;

[0117] Based on the mechanical properties of the deposited layer and the grain size, grain morphology and grain orientation distribution, an evaluation result of the influence of the heat input state on the microstructure and mechanical properties of the deposited layer is generated.

[0118] In practical applications, the acquired temperature field scanning data is first preprocessed. Specifically, the adaptive median filtering algorithm is used to remove thermal noise. The algorithm dynamically adjusts the filter window size and performs median filtering on each pixel, effectively removing thermal noise while retaining image details. For example, for a 32×32 pixel temperature field image, the initial window size can be set to 3×3, the maximum window size can be set to 7×7, and most of the thermal noise can be effectively removed by three iterations. Subsequently, the contrast-limited adaptive histogram equalization method is used to enhance the image contrast. This method divides the image into several sub-blocks, performs histogram equalization on each sub-block, and limits the contrast enhancement amplitude to avoid excessive noise enhancement. The image can be divided into 8×8 sub-blocks, and the contrast limit threshold is set to 0.01. After processing, the contrast of the temperature field image is significantly improved, which is conducive to subsequent analysis.

[0119] Next, the temperature gradient of the preprocessed temperature field data is calculated using a multi-scale gradient calculation method. Specifically, the gradients are calculated at three different spatial scales: pixel level, 3×3 neighborhood, and 5×5 neighborhood, and Gx1, Gx3, and Gx5 are obtained respectively. Then, the gradients of the three scales are weighted fused, and the weights can be set to 0.5, 0.3, and 0.2, and the total temperature gradient field G = 0.5Gx1 + 0.3Gx3 + 0.2Gx5 is obtained. This multi-scale fusion method not only retains the fine gradient information, but also takes into account the temperature change trend at a larger scale.

[0120] At the same time, the cooling rate of each pixel is calculated by analyzing the temperature changes between consecutive frames in the preprocessed temperature field data. Assuming that the time interval between two adjacent frames is 0.1 seconds, the cooling rate R = (T2 - T1) / 0.1, where T1 and T2 are the temperature values ​​of two adjacent frames, respectively. In order to smooth the cooling rate data, a Kalman filter is used for processing. The state transfer matrix of the Kalman filter can be set to the unit matrix, the observation matrix is ​​[1, 0], and the process noise covariance and observation noise covariance can be adjusted according to actual conditions, generally set to smaller values ​​such as 0.01 and 0.1. After filtering, smooth cooling rate data is obtained, which effectively suppresses the influence of instantaneous fluctuations.

[0121] Then, a long short-term memory network (LSTM) is used to establish a dynamic mapping relationship between the real-time heat input and the total temperature gradient field and the smoothed cooling rate data. The input of the LSTM network is the real-time heat input, and the output is the total temperature gradient field and cooling rate data. The network structure can be set to 2 layers of LSTM, each layer contains 64 neurons, and finally connected to a fully connected layer. The Adam optimizer is used, the learning rate is set to 0.001, and 500 epochs are trained. In this way, a heat input-temperature field feature mapping model is obtained, which can accurately predict the temperature field characteristics under different heat input conditions.

[0122] Using the above heat input-temperature field characteristic mapping model, based on the total temperature gradient field and the smoothed cooling rate data, a multiphase field model that couples the temperature field and concentration field is used to predict the grain characteristics of the deposited layer. The multiphase field model takes into account the coupling of the temperature field, concentration field and phase field, and simulates the grain growth process by solving a series of partial differential equations. Model parameters such as interface energy and kinetic coefficients can be adjusted according to the specific material. After simulation calculation, information such as the grain size, grain morphology and grain orientation distribution of the deposited layer can be obtained. For example, it can be predicted that the average grain size is 50μm, equiaxed crystals account for 70%, columnar crystals account for 30%, and the grain orientation is mainly.

[0123] Subsequently, the mechanical properties of the deposited layer are calculated using a multi-scale modeling method based on machine learning, using a pre-established microstructure-performance relationship database and the above-mentioned grain feature information. This method combines the crystal plasticity finite element model and the data-driven model, and can accurately predict the macroscopic mechanical properties under different microstructures. Specifically, a large amount of microstructure-performance sample data is first generated using a crystal plasticity finite element model, and then a machine learning model (such as a random forest or a neural network) is trained to establish a mapping relationship between microscopic features and macroscopic properties. Finally, the predicted grain features are input into the trained machine learning model to calculate the mechanical properties of the deposited layer, including yield strength, tensile strength, elongation, and hardness. For example, the yield strength is 450MPa, the tensile strength is 600MPa, the elongation is 15%, and the Vickers hardness is 200HV.

[0124] Finally, based on the calculated mechanical properties and grain characteristics of the deposited layer, the evaluation results of the influence of heat input state on the microstructure and mechanical properties of the deposited layer are generated. The evaluation results may include: 1) the relationship curve between heat input and grain size, such as when the heat input increases from 1kJ / mm to 3kJ / mm, the average grain size increases from 30μm to 80μm; 2) the relationship between cooling rate and grain morphology, such as when the cooling rate is greater than 100°C / s, equiaxed crystals are formed, and when it is less than 50°C / s, columnar crystals are formed; 3) the relationship between temperature gradient and grain orientation, such as when the temperature gradient is greater than 1000°C / mm, the grains tend to grow along the direction; 4) the corresponding relationship between microstructure characteristics and mechanical properties, such as when the grain size decreases by 10μm, the yield strength increases by about 20MPa. These quantitative relationships provide an important basis for optimizing heat input parameters and controlling the microstructure and properties of the deposited layer.

[0125] In this embodiment, adaptive median filtering and contrast-limited adaptive histogram equalization are used to remove thermal noise and enhance image contrast, ensuring that the temperature field data is clearer and more accurate; a multi-scale gradient method is used to calculate the total temperature gradient field, and the cooling rate data is smoothed by Kalman filtering to ensure the accuracy and stability of the gradient and rate and enhance the description of the heat conduction process; a long short-term memory network is used to establish a mapping relationship between heat input and temperature gradient and cooling rate, so as to achieve accurate modeling of the dynamic influence of heat input on the temperature field; the grain characteristics of the deposited layer are predicted by a multiphase field model, and the mechanical properties of the deposited layer are accurately estimated by combining the microstructure-performance database and machine learning technology, including key indicators such as yield strength and hardness; based on grain size, morphology and orientation, an evaluation of the impact of heat input state on microstructure and mechanical properties is generated to help optimize process parameters and improve forming quality.

[0126] In an optional implementation, generating a heat input sequence according to the impact relationship evaluation result includes:

[0127] Based on the results of the impact relationship assessment, the heat input is used as the decision variable, discretized into a time series, and the microstructure characteristics and mechanical properties are used as optimization targets to construct a multi-objective optimization model.

[0128] Using the Latin hypercube sampling method to generate an initial population in the decision space of the multi-objective optimization model, and using a deep neural network to construct a mapping relationship from a hot input sequence to the optimization objective to form a proxy model;

[0129] According to a preset population division ratio, the initial population is randomly divided into a first subgroup and a second subgroup, and computing resources are allocated accordingly, including: using a heat input-temperature field feature mapping model and a multiphase field model to perform accuracy evaluation on individuals in the first subgroup, and using the proxy model to perform rapid evaluation on individuals in the second subgroup; determining an objective function value for each individual in the initial population, and based on the obtained objective function value, using a fast non-dominated sorting algorithm to stratify the initial population, obtaining stratification results, and calculating the crowding distance of each individual;

[0130] According to the stratification result and crowding distance of the initial population, in combination with the current iteration stage, the parameters of the evolutionary operator are dynamically adjusted, and the alternative probabilities corresponding to the preset multiple evolutionary operators are updated by an adaptive strategy, and the evolutionary operator is determined by a roulette selection algorithm; based on the tournament selection method, the parent individual is selected from the initial population using the stratification result and crowding distance, and the determined evolutionary operator is applied to generate the offspring individual, the offspring individual is evaluated to obtain the objective function value corresponding to each offspring individual, the parent individual and the offspring individual are merged to form a merged population, non-dominated sorting is performed based on the objective function value of each individual in the merged population, and the next generation population is selected from the merged population in combination with the crowding distance; based on the next generation population, in the target space, a point inferior to all non-dominated solutions is selected as a reference point, the exclusive hypervolume of each individual is calculated, the hypervolume contribution of each individual is determined, and the computing resources are allocated according to the hypervolume contribution;

[0131] The iteration is repeated until the preset total number of iterations is reached to obtain the final non-dominated solution set, and the TOPSIS method is applied to select the best compromise solution based on the objective function value of each solution in the final non-dominated solution set to obtain the optimal heat input sequence.

[0132] The Latin hypercube sampling method specifically refers to a random sampling technique for multidimensional space, which can generate representative and evenly distributed sample sets. The basic idea is to divide the value range of each variable into multiple intervals of equal probability, and randomly select a value in each interval to ensure that the value of each dimension is fully covered. This method is widely used in high-dimensional optimization problems, and can maintain good coverage while reducing the number of samples;

[0133] The reference point specifically refers to an artificially defined ideal point in multi-objective optimization, which is used to evaluate the deviation of the solution set relative to a certain goal. The reference point usually does not belong to the actual solution space, but is used to measure the quality of the solution set, such as calculating the performance of an individual when searching for a solution in the non-dominated solution set. The reference point can help guide the search direction in the target space and promote the optimization solution set to converge towards a better objective function value;

[0134] The exclusive hypervolume specifically refers to an indicator that measures the contribution of a solution in multi-objective optimization. It represents the volume of the area formed by the solution and the reference point in the target space. Specifically, it is the exclusive part of the solution in the target space, which is not covered by other solutions. The larger the exclusive hypervolume, the greater the contribution of the solution to the entire solution set;

[0135] The hypervolume contribution specifically refers to the measure of the importance of a solution to the overall non-dominated solution set. It refers to the ratio of the exclusive hypervolume of a solution to the total hypervolume of all solutions in the entire non-dominated solution set. The larger the hypervolume contribution, the greater the contribution of the solution to the solution set. In evolutionary algorithms, such solutions often have higher retention value.

[0136] Firstly, based on the results of the impact relationship assessment, the heat input is used as the decision variable, which is discretized into a time series, and the microstructure characteristics and mechanical properties are used as optimization targets to construct a multi-objective optimization model. The decision variable of this model is the heat input at different times, and the objective function includes microstructure and performance indicators such as grain size, phase fraction, hardness, and strength.

[0137] Next, the Latin hypercube sampling method is used to generate the initial population in the decision space. Latin hypercube sampling can ensure that the samples are evenly distributed in all dimensions and improve sampling efficiency. For example, assuming that the decision variable is the heat input at 5 moments, the value range of each variable is 0-1000J, and the population size is 100, then 100 evenly distributed sample points can be generated in the space of [0, 1000]^5. Then, a deep neural network is used to construct a mapping relationship from the heat input sequence to the optimization target to form a proxy model. The neural network takes the heat input sequence as input and the microstructure and performance indicators as output. It is trained through a large amount of simulated data and can quickly predict the objective function value corresponding to any heat input sequence.

[0138] Then, the initial population is randomly divided into the first subgroup and the second subgroup according to the preset population division ratio (such as 7:3), and computing resources are allocated. The individuals in the first subgroup are accurately evaluated using the heat input-temperature field feature mapping model and the multiphase field model to calculate their true objective function value. The individuals in the second subgroup are quickly evaluated using the proxy model to obtain an approximate objective function value. After determining the objective function value of each individual in the initial population, the initial population is stratified using the fast non-dominated sorting algorithm to obtain the stratification result. At the same time, the crowding distance of each individual is calculated to distinguish the pros and cons of individuals in the same layer.

[0139] According to the stratification results and crowding distance of the initial population, combined with the current iteration stage, the parameters of the evolutionary operator are dynamically adjusted. For example, a larger mutation rate is used in the early stage to increase population diversity, and the mutation rate is reduced in the later stage to accelerate convergence. At the same time, an adaptive strategy is used to update the alternative probabilities corresponding to multiple preset evolutionary operators (such as simulated binary crossover, polynomial mutation, etc.), and the roulette wheel selection algorithm is used to determine which operator to use. Then, based on the tournament selection method, the parent individuals are selected from the initial population using the stratification results and crowding distance, and the determined evolutionary operator is applied to generate the offspring individuals. The offspring individuals are evaluated to obtain the objective function value corresponding to each offspring individual. The parent individuals and offspring individuals are merged to form a merged population, and a non-dominated sort is performed based on the objective function value of each individual in the merged population, and the next generation population is selected from the merged population in combination with the crowding distance.

[0140] Based on the next generation population, a point inferior to all non-dominated solutions is selected as a reference point in the objective space, and the exclusive hypervolume of each individual is calculated to determine the hypervolume contribution of each individual. For example, assuming there are 3 objective functions, the reference point coordinates are (1, 1, 1), and the objective function value of an individual is (0.2, 0.3, 0.4), then its exclusive hypervolume is (1-0.2)*(1-0.3)*(1-0.4)=0.336. According to the hypervolume contribution, computing resources are allocated, and more resources are allocated to individuals with large contributions for accurate evaluation.

[0141] Repeat the above iterative process until the preset total number of iterations (such as 100 generations) is reached to obtain the final non-dominated solution set. The TOPSIS method is applied to select the best compromise solution based on the objective function value of each solution in the final non-dominated solution set. Specifically, the positive ideal solution and the negative ideal solution are first determined, and then the distance from each solution to the positive and negative ideal solutions is calculated. The solution closest to the positive ideal solution and farthest from the negative ideal solution is selected as the best compromise solution, and the corresponding heat input sequence is the optimal heat input sequence.

[0142] Through the above method, the optimal heat input sequence considering multiple objectives can be effectively generated to provide guidance for welding process optimization. This method combines the exact model and the proxy model to improve the computational efficiency; adopts multiple evolutionary strategies to enhance the algorithm's search capability; introduces the hypervolume indicator for resource allocation to improve the quality of the Pareto frontier.

[0143] In this embodiment, an effective multi-objective optimization model is established by discretizing the heat input into a time series and taking the microstructure characteristics and mechanical properties as optimization objectives, which provides a scientific basis for decision-making; the Latin hypercube sampling method is used to generate the initial population to ensure that a fully representative sample is obtained in the decision space, thereby improving the optimization efficiency; a mapping relationship between the heat input sequence and the optimization objective is constructed through a deep neural network to form an agent model, thereby achieving rapid evaluation and accurate prediction and reducing the computational cost; during the optimization process, the parameters of the evolutionary operator are dynamically adjusted according to the stratification results and the crowding distance, thereby improving the flexibility and efficiency of the search strategy; the fast non-dominated sorting algorithm and the crowding distance are used to evaluate individuals, thereby ensuring the diversity of the population and the screening of high-quality solutions, which is helpful to find the Pareto optimal solution set; the computing resources are dynamically allocated by calculating the hypervolume contribution, thereby improving the efficiency of the optimization process and ensuring the rapid finding of the optimal solution; finally, the best compromise solution is selected from the non-dominated solution set through the TOPSIS method, thereby providing a practical heat input sequence, which is helpful to improve the microstructure and mechanical properties of the deposited layer.

[0144] In an optional embodiment, the heat input sequence is input into an adaptive control system constructed based on a reinforcement learning algorithm, and the optimal parameter scheme is calculated according to the preset target deposition layer performance index and the current temperature field state, including:

[0145] The heat input sequence is discretized and normalized at fixed time intervals, and a sliding window is used to treat multiple consecutive time steps as an input unit;

[0146] Use a three-dimensional convolutional neural network to compress the temperature field information, convert the target performance index into a vector representation, encode the current layer geometric characteristics and material thermophysical parameters, and construct the state space;

[0147] Based on the state space, a hybrid action space is determined, including a discrete action space and a continuous action space, wherein the discrete action space includes current, voltage, wire feeding speed, pulse frequency and duty cycle, and the continuous action space is used to fine-grainedly adjust corresponding parameters in the discrete action space;

[0148] Construct a multi-objective reward function, including quality reward, geometric accuracy reward, energy efficiency reward and stability reward, and use weight coefficients to balance the reward function, where the quality reward is calculated based on the deviation of hardness and strength from the target value, the geometric accuracy reward is determined based on the evaluation of shape accuracy and surface roughness, the energy efficiency reward evaluates the heat input utilization and energy loss, and the stability reward is based on arc stability and spatter degree;

[0149] Initialize the main Q network and the target Q network, wherein the main Q network and the target Q network have the same network structure and different initial parameter values; design a dual-path deep Q network architecture, and divide the output layers of the main Q network and the target Q network into a state value function branch and an advantage function branch, wherein the state value function branch estimates the state value, and the advantage function branch estimates the action advantage; construct an experience playback buffer to store experience samples of the interaction between the agent and the environment, wherein each experience sample contains the current state, the executed action, the reward obtained, the next state, and the termination flag;

[0150] Implement a multi-step bootstrapping mechanism to calculate the target Q value by accumulating discounted rewards over multiple steps and using the target Q network to estimate the state value after multiple steps;

[0151] defining a loss function that combines a temporal difference error and a dual-path architecture to calculate a difference between a Q value predicted by the main Q network and a target Q value;

[0152] Using the samples in the experience replay buffer, calculating the value of the loss function, and updating the parameters of the main Q network based on the value of the loss function;

[0153] Copying the parameters of the master Q network to the target Q network according to a predefined update frequency;

[0154] Repeat until the number of training times reaches the preset value;

[0155] Using the soft actor-critic algorithm, based on the trained reinforcement learning model, the current state is input, including the processed thermal input sequence, temperature field information and target performance indicators, the trained Q network is used to evaluate the value of different actions, and the optimal action is selected through soft maximization operation;

[0156] Through multiple iterations of optimization, the optimal parameter solution is calculated.

[0157] In a specific implementation, the heat input sequence is first discretized at fixed time intervals (e.g., 0.1 seconds), and the values ​​are normalized so that they fall within the interval [0, 1]. Then, a sliding window technique is used to treat multiple consecutive time steps (e.g., 10) as an input unit to form an input sequence. For example, for an original heat input sequence of length 100, 91 10-dimensional input units can be obtained after processing.

[0158] Next, a three-dimensional convolutional neural network is used to compress the temperature field information. The network structure contains three three-dimensional convolutional layers, each followed by BatchNormalization and ReLU activation functions, and finally a fully connected layer is used to map the features to low-dimensional vectors. The target performance indicators (such as hardness, strength, etc.) are converted into vector representations through the fully connected layer. The geometric features of the current layer (such as layer height, width, etc.) and the material thermophysical parameters (such as thermal conductivity, specific heat capacity, etc.) are also encoded. These vectors are spliced ​​together to form a state space.

[0159] Based on the state space, the hybrid action space is determined. The discrete action space includes current (such as 100A, 150A, 200A), voltage (such as 20V, 25V, 30V), wire feeding speed (such as 5m / min, 7m / min, 9m / min), pulse frequency (such as 50Hz, 100Hz, 150Hz) and duty cycle (such as 30%, 50%, 70%). The continuous action space is used for fine-grained adjustment near discrete values, such as the current can be continuously adjusted within the range of ±10A.

[0160] Construct a multi-objective reward function, including quality reward, geometric accuracy reward, energy efficiency reward, and stability reward. The quality reward is calculated based on the deviation of hardness and strength from the target value. The smaller the deviation, the higher the reward. The geometric accuracy reward is determined by evaluating the shape accuracy and surface roughness. The higher the accuracy, the greater the reward. The energy efficiency reward evaluates the heat input utilization and energy loss. The higher the efficiency, the greater the reward. The stability reward is based on arc stability and spatter. The better the stability, the higher the reward. Use weight coefficients (such as 0.3, 0.3, 0.2, 0.2) to balance each reward.

[0161] Initialize the main Q network and the target Q network. The two networks have the same structure but different initial parameters. The network structure adopts a dual-path deep Q network architecture, and the output layer is divided into a state value function branch and an advantage function branch. The state value function branch estimates the overall value of the state, and the advantage function branch estimates the advantage of each action relative to the average level.

[0162] Construct an experience replay buffer to store experience samples of the agent's interaction with the environment. Each sample contains the current state, the action performed, the reward obtained, the next state, and the termination flag. The buffer size is set to 10,000 and is updated in a first-in, first-out manner.

[0163] Implement a multi-step bootstrapping mechanism, accumulate discounted rewards for 3 steps and use the target Q network to estimate the state value 3 steps later to calculate the target Q value. The discount factor is set to 0.99.

[0164] Define the loss function, combine the temporal difference error and the two-way architecture, and calculate the difference between the Q value predicted by the main Q network and the target Q value. Each training run randomly samples 32 samples from the experience replay buffer to calculate the loss.

[0165] The main Q network parameters are updated using the Adam optimizer with a learning rate of 0.001. The main Q network parameters are copied to the target Q network every 100 steps. The training process is repeated until 100,000 steps are reached.

[0166] Use the soft actor-critic algorithm to make decisions based on the trained reinforcement learning model. Input the current state, including the processed thermal input sequence, temperature field information, and target performance indicators. Use the trained Q network to evaluate the value of different actions, and calculate the selection probability of each action through a softmax operation, with the temperature parameter set to 0.1. Randomly select actions based on the probability distribution.

[0167] Through multiple iterations of optimization (e.g., 1000 times), the status and execution actions are continuously updated, rewards are accumulated, and parameter solutions are recorded. Finally, the parameter solution with the highest cumulative reward is selected as the optimal solution output. For example, the optimal solution may be: current 180A, voltage 27V, wire feeding speed 8m / min, pulse frequency 120Hz, duty cycle 60%.

[0168] This method adaptively optimizes arc additive manufacturing process parameters through reinforcement learning, and can automatically adjust process parameters to obtain the best deposition effect based on the real-time temperature field and target performance indicators.

[0169] In this embodiment, by discretizing and normalizing the heat input sequence and adopting the sliding window technology, it is helpful to more accurately capture and analyze the heat input changes in a dynamic environment, thereby improving the accuracy of the subsequent model; by using a three-dimensional convolutional neural network to compress the temperature field information and convert the performance indicators into vector representations, it is possible to effectively extract the geometric characteristics of the current layer and the thermal physical parameters of the material, construct a comprehensive state space, and enhance the expressive power of the model; a clear distinction is made between discrete and continuous action spaces, so that higher flexibility and fine-grained control are provided when adjusting the welding parameters, and the adjustment strategy of the process parameters is optimized; by constructing a multi-objective reward function, multiple aspects such as quality, geometric accuracy, energy efficiency and stability are comprehensively considered, so that the model The training of the model is more comprehensive, thereby improving the quality and performance of the final deposited layer; the learning ability and prediction accuracy of the model are enhanced through the branching structure of the state value function and the advantage function, so that the intelligent agent can better evaluate the potential value of different actions; the experience replay buffer is used to store the experience samples of the interaction between the intelligent agent and the environment, combined with the multi-step bootstrap mechanism, the stability and efficiency of the learning process are improved, and the risk of overfitting is reduced; by setting the update frequency and parameter replication mechanism, the effective coordination between the target Q network and the main Q network is ensured, and the convergence of the model is enhanced; finally, the optimal action is selected by the reinforcement learning model, and multiple iterations of optimization are performed to effectively obtain the optimal welding parameter solution and improve the consistency of the organization and performance of the deposited layer.

[0170] In an optional embodiment, implementing a multi-step bootstrapping mechanism, calculating a target Q value by accumulating discounted rewards for multiple steps and using the target Q network to estimate the state value after multiple steps includes:

[0171] Set a step number parameter to determine the number of time steps to consider when calculating the target Q value; collect and store state, action, reward, and next state information at each time step until continuous experience that meets the step number parameter is collected;

[0172] For the current time step, calculate a multi-step reward value, which includes all rewards from the current time step to the future step number parameter minus one time step, and the estimated Q value of the last time step, where the corresponding reward of each time step is multiplied by the corresponding discount factor;

[0173] Based on the multi-step reward values, a target Q value is determined; a difference between the target Q value and the current Q value is calculated to determine a time difference error, and based on the time difference error, a gradient descent method is used to minimize the mean square error loss and update the main Q network parameters;

[0174] The sliding window method is used to continuously update multi-step experience. At the beginning of a new time step, the earliest experience in the experience replay buffer is removed and the new experience is added to the experience replay buffer.

[0175] In the implementation of the multi-step bootstrapping mechanism, you first need to set a step number parameter N, which determines the number of time steps forward considered when calculating the target Q value. Usually N can be set to an integer value between 3 and 5, for example N=4. This means that when calculating the target Q value, the information of the current time step and the next 3 time steps will be considered.

[0176] Next, at each time step t, the agent interacts with the environment and collects experience data, including the current state st, the action performed at, the reward rt, and the next state st+1 transferred to. This information is stored in the experience replay buffer until the experience of N consecutive time steps is accumulated. For example, assuming N=4, it is necessary to collect the experience of 4 consecutive time steps (st, at, rt, st+1), (st+1, at+1, rt+1, st+2), (st+2, at+2, rt+2, st+3), and (st+3, at+3, rt+3, st+4).

[0177] When enough continuous experience is collected, the multi-step reward value Gt can be calculated. The multi-step reward value contains all rewards from the current time step t to the t+N-1 time step, as well as the state value estimated by the target Q network at the t+N time step. The specific calculation process is as follows:

[0178] First, for time steps t to t+N-1, multiply the reward rk at each time step k by the discount factor γ raised to the power of kt. For example, assuming γ = 0.9, then rt remains unchanged, rt+1 is multiplied by 0.9, rt+2 is multiplied by 0.81, and rt+3 is multiplied by 0.729.

[0179] Then, the target Q network is used to estimate the maximum Q value of the state st+N at time step t+N, recorded as maxQ(st+N,a). This estimate is multiplied by γ to the power of N.

[0180] Finally, add all the above items to get the multi-step reward value Gt:

[0181] Gt = rt + 0.9rt+1 + 0.81rt+2 + 0.729rt+3 + 0.6561maxQ(st+4, a);

[0182] Based on the calculated multi-step reward value Gt, the target Q value ytarget can be determined. For the state-action pair (st, at) at time step t, its target Q value is equal to the multi-step reward value Gt.

[0183] Next, the difference between the target Q value ytarget and the Q value Q(st, at) estimated by the current Q network is calculated to obtain the temporal difference error δt:

[0184] δt = ytarget - Q(st,at);

[0185] Based on the calculated time difference error δt, the mean square error loss is minimized using the gradient descent method to update the parameters θ of the main Q network. Specifically, the following parameter update rule can be used:

[0186] ;

[0187] Where α is the learning rate, represents the gradient of the Q value with respect to the parameter θ.

[0188] After completing a parameter update, the sliding window method is used to continuously update the multi-step experience. Specifically, at the beginning of the new time step t+1, the earliest experience (st-N+1, at-N+1, rt-N+1, st-N+2) in the experience playback buffer is removed, and the newly collected experience (st+1, at+1, rt+1, st+2) is added to the buffer. In this way, the buffer can always contain the experience of the most recent N consecutive time steps.

[0189] Through the above steps, a multi-step bootstrapping mechanism is implemented, which can effectively use multi-step reward information to calculate a more accurate target Q value, thereby improving the performance of the reinforcement learning algorithm. This method combines the advantages of TD(λ) and Q-learning, taking into account long-term rewards while retaining the flexibility of off-policy learning.

[0190] In this embodiment, by setting the step number parameter, the time range that needs to be considered forward when calculating the target Q value is clarified, which helps the model to better evaluate future rewards and improve the foresight of decision-making; by calculating multi-step reward values, all reward information from the current time step to the future time step is integrated to ensure that the calculation of the Q value is more comprehensive and accurate, thereby optimizing the learning process; based on the difference between the target Q value and the current Q value, the time difference error is calculated, which helps to understand the deviation of the model prediction and further guide the parameter update of the model; by minimizing the mean square error loss using the gradient descent method, the effective update of the main Q network parameters is ensured, thereby improving the learning efficiency and convergence speed; the sliding window method is used to continuously update the multi-step experience, so that the model always learns based on the latest experience, thereby improving the model's adaptability to environmental changes and real-time performance.

[0191] Figure 2 Schematic diagram of the structure of the heat input control system of the arc additive continuous forming according to the embodiment of the present invention. Figure 2 As shown, the system comprises:

[0192] The first unit is used to collect process parameters in the arc additive continuous forming process in real time through multiple sensors, obtain pre-processing process data by filtering and standardizing the process parameters; input the pre-processing process data into an adaptive heat input calculation model, the adaptive heat input calculation model is constructed based on thermodynamic principles and heat transfer theory, and calculates the real-time heat input in the deposition process;

[0193] The second unit is used to start the multispectral imaging device based on the real-time heat input, perform real-time temperature field scanning on the surface of the sediment layer, record the temperature field scanning data, extract the temperature gradient and cooling rate information of each region of the sediment layer through image processing algorithm analysis, associate the temperature gradient and cooling rate information with the real-time heat input, and determine the impact relationship evaluation result of the current heat input state on the microstructure and mechanical properties of the sediment layer; generate a heat input sequence according to the impact relationship evaluation result;

[0194] The third unit is used to input the heat input sequence into the adaptive control system built based on the reinforcement learning algorithm, and calculate the optimal parameter scheme according to the preset target deposition layer performance indicators and the current temperature field state; based on the optimal parameter scheme, the process parameters are adjusted in real time through the digital welding power supply and the high-precision wire feeding system, and the arc pulse modulation device is started at the same time. The corresponding time and space distribution of the control energy input is adjusted according to the optimal parameter scheme, and the heat input of different positions and different layers of the deposition layer is controlled based on the time and space distribution.

[0195] According to a third aspect of the embodiments of the present invention,

[0196] An electronic device is provided, comprising:

[0197] processor;

[0198] a memory for storing processor-executable instructions;

[0199] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0200] A fourth aspect of the embodiments of the present invention is:

[0201] 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.

[0202] 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.

[0203] 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 heat input in arc additive continuous forming, characterized in that: include: The process parameters in the arc additive continuous forming process are collected in real time by multiple sensors, and the pre-processing process data are obtained by filtering and standardizing the process parameters; Inputting the pretreatment process data into an adaptive heat input calculation model, wherein the adaptive heat input calculation model is constructed based on thermodynamic principles and heat transfer theory to calculate the real-time heat input during the deposition process; Based on the real-time heat input, a multispectral imaging device is started to perform a real-time temperature field scan on the surface of the deposited layer, and the temperature field scan data is recorded. The temperature gradient and cooling rate information of each region of the deposited layer are extracted through image processing algorithm analysis, and the temperature gradient and cooling rate information are correlated with the real-time heat input to determine the evaluation result of the relationship between the current heat input state and the microstructure and mechanical properties of the deposited layer; Generate a heat input sequence based on the impact relationship evaluation results; The heat input sequence is input into the adaptive control system built based on the reinforcement learning algorithm, and the optimal parameter scheme is calculated according to the preset target deposition layer performance index and the current temperature field state; based on the optimal parameter scheme, the process parameters are adjusted in real time through the digital welding power supply and the high-precision wire feeding system, and the arc pulse modulation device is started at the same time, and the corresponding time and space distribution of the control energy input is adjusted according to the optimal parameter scheme. Based on the time and space distribution, the heat input of different positions and different layers of the deposition layer is controlled; The real-time heat input during the deposition process is calculated as follows: Construct multi-scale nodes, define microstructure features as microscale nodes, cross-sectional grid points as mesoscale nodes, and the centers of surface partitioned areas as macroscale nodes; Based on the heat conduction theory, the spatial connection between multi-scale nodes is established through physical distance; based on different time steps, the temporal connection between multi-scale nodes is established; based on each scale in the multi-scale nodes, cross-scale connections are established to form a multi-scale space-time graph; Performing a crystal graph convolution operation on the micro-scale nodes, performing an anisotropic graph convolution operation on the meso-scale nodes, and performing a spherical graph convolution operation on the macro-scale nodes to obtain multi-scale features; bidirectionally transmitting the multi-scale features through upsampling and downsampling operations, adjusting the scale importance through an attention mechanism, and performing scale fusion to obtain fused multi-scale features; Processing the fused multi-scale features through a graph converter to obtain long-term temporal dependency information, inputting the long-term temporal dependency information into a graph differential equation, simulating a dynamic process of heat conduction, and obtaining a preliminary heat input prediction result; Based on the preliminary heat input prediction result, calculate the material phase change latent heat compensation, convert the material constitutive equation into a Neural Ordinary Differential Equation, use the differentiable yield criterion to calculate the material property calculation result, combine the above material property calculation result with the preliminary heat input prediction result, and obtain a revised heat input prediction result; Based on the corrected heat input prediction results, a loss function including energy conservation constraints, thermodynamic law constraints and difficult sample loss terms is constructed, a gradient penalty term is added, and the weights of the corresponding terms in the loss function are dynamically adjusted based on the different stages of the arc additive forming process. Based on the goal of minimizing the loss function, the adaptive heat input calculation model is iteratively trained to output the final real-time heat input prediction result.

2. The method according to claim 1, characterized in that Through image processing algorithm analysis, the temperature gradient and cooling rate information of each area of ​​the deposit layer are extracted, and the temperature gradient and cooling rate information are correlated with the real-time heat input to determine the impact of the current heat input state on the microstructure and mechanical properties of the deposit layer. The evaluation results include: Performing image preprocessing on the temperature field scanning data, including using an adaptive median filtering algorithm to remove thermal noise, and using a contrast-limited adaptive histogram equalization method to enhance image contrast, to obtain preprocessed temperature field data; The temperature gradient of the preprocessed temperature field data is calculated using a multi-scale gradient, and the gradient is calculated at different spatial scales and weighted fusion is performed to obtain a total temperature gradient field; the cooling rate of each pixel is calculated by analyzing the temperature change between consecutive frames in the preprocessed temperature field data, and the cooling rate is smoothed using a Kalman filter to obtain smoothed cooling rate data; Using a long short-term memory network to establish a dynamic mapping relationship between the real-time heat input and the total temperature gradient field and the smoothed cooling rate data, a heat input-temperature field characteristic mapping model is obtained; The heat input-temperature field characteristic mapping model is used, based on the total temperature gradient field and the smoothed cooling rate data, and a multiphase field model of coupled temperature field and concentration field is used to predict the grain size, grain morphology and grain orientation distribution of the deposited layer; the mechanical properties of the deposited layer, including yield strength, tensile strength, elongation and hardness, are calculated using a multiscale modeling method based on machine learning, in combination with the grain size, grain morphology and grain orientation distribution, using a pre-established microstructure-performance relationship database; Based on the mechanical properties of the deposited layer and the grain size, grain morphology and grain orientation distribution, an evaluation result of the influence relationship of the heat input state on the microstructure and mechanical properties of the deposited layer is generated.

3. The method according to claim 2, characterized in that Based on the impact relationship evaluation results, the heat input sequence is generated including: Based on the results of the impact relationship assessment, the heat input is used as the decision variable, discretized into a time series, and the microstructure characteristics and mechanical properties are used as optimization targets to construct a multi-objective optimization model. Using the Latin hypercube sampling method to generate an initial population in the decision space of the multi-objective optimization model, and using a deep neural network to construct a mapping relationship from a hot input sequence to the optimization objective to form a proxy model; According to a preset population division ratio, the initial population is randomly divided into a first subgroup and a second subgroup, and computing resources are allocated accordingly, including: using a heat input-temperature field feature mapping model and a multiphase field model to perform accuracy evaluation on individuals in the first subgroup, and using the proxy model to perform rapid evaluation on individuals in the second subgroup; determining an objective function value for each individual in the initial population, and based on the obtained objective function value, using a fast non-dominated sorting algorithm to stratify the initial population, obtaining stratification results, and calculating the crowding distance of each individual; According to the stratification result and crowding distance of the initial population, in combination with the current iteration stage, the parameters of the evolutionary operator are dynamically adjusted, and the alternative probabilities corresponding to the preset multiple evolutionary operators are updated by an adaptive strategy, and the evolutionary operator is determined by a roulette selection algorithm; based on the tournament selection method, the parent individual is selected from the initial population using the stratification result and crowding distance, and the determined evolutionary operator is applied to generate the offspring individual, the offspring individual is evaluated to obtain the objective function value corresponding to each offspring individual, the parent individual and the offspring individual are merged to form a merged population, non-dominated sorting is performed based on the objective function value of each individual in the merged population, and the next generation population is selected from the merged population in combination with the crowding distance; based on the next generation population, in the target space, a point inferior to all non-dominated solutions is selected as a reference point, the exclusive hypervolume of each individual is calculated, the hypervolume contribution of each individual is determined, and the computing resources are allocated according to the hypervolume contribution; The iteration is repeated until the preset total number of iterations is reached to obtain the final non-dominated solution set, and the TOPSIS method is applied to select the best compromise solution based on the objective function value of each solution in the final non-dominated solution set to obtain the optimal heat input sequence.

4. The method according to claim 1, characterized in that: The heat input sequence is input into the adaptive control system built based on the reinforcement learning algorithm. According to the preset target deposition layer performance index and the current temperature field state, the optimal parameter scheme is calculated, including: The heat input sequence is discretized and normalized at fixed time intervals, and a sliding window is used to treat multiple consecutive time steps as an input unit; Use a three-dimensional convolutional neural network to compress the temperature field information, convert the target performance index into a vector representation, encode the current layer geometric characteristics and material thermophysical parameters, and construct the state space; Based on the state space, a hybrid action space is determined, including a discrete action space and a continuous action space, wherein the discrete action space includes current, voltage, wire feeding speed, pulse frequency and duty cycle, and the continuous action space is used to fine-grainedly adjust corresponding parameters in the discrete action space; Construct a multi-objective reward function, including quality reward, geometric accuracy reward, energy efficiency reward and stability reward, and use weight coefficients to balance the reward function, where the quality reward is calculated based on the deviation of hardness and strength from the target value, the geometric accuracy reward is determined based on the evaluation of shape accuracy and surface roughness, the energy efficiency reward evaluates the heat input utilization and energy loss, and the stability reward is based on arc stability and spatter degree; Initialize the main Q network and the target Q network, wherein the main Q network and the target Q network have the same network structure and different initial parameter values; design a dual-path deep Q network architecture, and divide the output layers of the main Q network and the target Q network into a state value function branch and an advantage function branch, wherein the state value function branch estimates the state value, and the advantage function branch estimates the action advantage; construct an experience playback buffer to store experience samples of the interaction between the agent and the environment, wherein each experience sample contains the current state, the executed action, the reward obtained, the next state, and the termination flag; Implement a multi-step bootstrapping mechanism to calculate the target Q value by accumulating discounted rewards over multiple steps and using the target Q network to estimate the state value after multiple steps; defining a loss function that combines a temporal difference error and a dual-path architecture to calculate a difference between a Q value predicted by the main Q network and a target Q value; Using the samples in the experience replay buffer, calculating the value of the loss function, and updating the parameters of the main Q network based on the value of the loss function; Copying the parameters of the master Q network to the target Q network according to a predefined update frequency; Repeat until the number of training times reaches the preset value; Using the soft actor-critic algorithm, based on the trained reinforcement learning model, the current state is input, including the processed thermal input sequence, temperature field information and target performance indicators, the trained Q network is used to evaluate the value of different actions, and the optimal action is selected through soft maximization operation; Through multiple iterations of optimization, the optimal parameter solution is calculated.

5. The method according to claim 4, characterized in that Implementing a multi-step bootstrapping mechanism to calculate the target Q value by accumulating discounted rewards over multiple steps and using the target Q network to estimate the state value after multiple steps includes: Set a step number parameter to determine the number of time steps to consider when calculating the target Q value; collect and store state, action, reward, and next state information at each time step until continuous experience that meets the step number parameter is collected; For the current time step, calculate a multi-step reward value, which includes all rewards from the current time step to the future step number parameter minus one time step, and the estimated Q value of the last time step, where the corresponding reward of each time step is multiplied by the corresponding discount factor; Based on the multi-step reward values, a target Q value is determined; a difference between the target Q value and the current Q value is calculated to determine a time difference error, and based on the time difference error, a gradient descent method is used to minimize the mean square error loss and update the main Q network parameters; The sliding window method is used to continuously update multi-step experience. At the beginning of a new time step, the earliest experience in the experience replay buffer is removed and the new experience is added to the experience replay buffer.

6. A heat input control system for arc additive continuous forming, used to implement the method described in any one of claims 1 to 5, characterized in that: include: The first unit is used to collect process parameters in the arc additive continuous forming process in real time through multiple sensors, and obtain pre-processing process data by filtering and standardizing the process parameters; Inputting the pretreatment process data into an adaptive heat input calculation model, wherein the adaptive heat input calculation model is constructed based on thermodynamic principles and heat transfer theory to calculate the real-time heat input during the deposition process; The second unit is used to start the multispectral imaging device based on the real-time heat input, perform real-time temperature field scanning on the surface of the deposited layer, record the temperature field scanning data, extract the temperature gradient and cooling rate information of each area of ​​the deposited layer through image processing algorithm analysis, and associate the temperature gradient and cooling rate information with the real-time heat input to determine the evaluation result of the influence of the current heat input state on the microstructure and mechanical properties of the deposited layer; Generate a heat input sequence based on the impact relationship evaluation results; The third unit is used to input the heat input sequence into the adaptive control system built based on the reinforcement learning algorithm, and calculate the optimal parameter scheme according to the preset target deposition layer performance indicators and the current temperature field state; based on the optimal parameter scheme, the process parameters are adjusted in real time through the digital welding power supply and the high-precision wire feeding system, and the arc pulse modulation device is started at the same time. The corresponding time and space distribution of the control energy input is adjusted according to the optimal parameter scheme, and the heat input of different positions and different layers of the deposition layer is controlled based on the time and space distribution.

7. 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 5.

8. 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 5 is implemented.

Images