Steel-aluminum dissimilar metal additive manufacturing interface regulation method and device

Abstract

This invention provides a method and apparatus for interface control in steel-aluminum dissimilar metal additive manufacturing, relating to the field of additive manufacturing technology. The method mainly includes: acquiring current process parameters and performing additive manufacturing based on these parameters; acquiring interface temperature field data during the manufacturing process; determining temperature characteristic parameters based on the interface temperature field data; inputting the temperature characteristic parameters into a prediction model to obtain interface brittle phase characteristics; determining the interface bonding strength based on the interface brittle phase characteristics; responding to the interface bonding strength not meeting a preset threshold, calling a reverse adjustment model to determine the process parameter adjustment amount; and adjusting the additive manufacturing process parameters based on the process parameter adjustment amount. This invention enables proactive and precise control of the morphology and layer thickness of the interface brittle phase, thereby improving the interface bonding strength.

Description

Technical Field

[0001] This invention relates to the field of electric arc additive manufacturing technology, and more specifically, to a method and apparatus for interface control in steel-aluminum dissimilar metal additive manufacturing. Background Technology

[0002] With the increasing demand for lightweighting in new energy vehicles, steel-aluminum composite structures, combining the high strength of steel with the low density of aluminum alloys, have been widely used in automobile body manufacturing. Arc additive manufacturing technology, with its high forming efficiency, high material utilization, and adaptability to large-size components, has become a core technology for achieving integrated forming of steel-aluminum composite components.

[0003] However, due to the significant differences in thermophysical properties and poor miscibility between steel and aluminum, hard and brittle intermetallic compounds are easily formed at the interface during additive manufacturing. These brittle phases are usually tongue-shaped or needle-shaped, which can easily cause stress concentration, resulting in low interfacial bonding strength, easy cracking, and seriously affecting the quality of the component.

[0004] Existing technologies typically employ open-loop parameter control, which cannot sense interface temperature changes in real time or predict the evolution of brittle phases. This results in blind and delayed adjustments to process parameters, making it difficult to achieve precise control over interface quality. Therefore, there is an urgent need for an intelligent method capable of real-time sensing, prediction, and closed-loop regulation of interface quality. Summary of the Invention

[0005] This invention provides a method and apparatus for interface control in steel-aluminum dissimilar metal additive manufacturing. It addresses the problems of uncontrolled morphology of brittle phases at the interface, low interfacial bonding strength, and lagging process parameter control in existing steel-aluminum additive manufacturing processes. This invention primarily utilizes a solid-liquid interface additive manufacturing layout with steel preceding aluminum, combined with a three-level cascaded neural network performance prediction and parameter adjustment inverse model to achieve a complete prediction-adjustment closed loop. This induces the brittle phase to exhibit a spherical FeAl3 morphology, enabling proactive and precise control of the morphology and layer thickness of the brittle phase at the interface, thereby improving the interfacial bonding strength.

[0006] The technical means employed in this invention are as follows:

[0007] A method for controlling the interface in steel-aluminum dissimilar metal additive manufacturing is applied to a dual-gun electric arc additive manufacturing system, wherein the dual-gun electric arc additive manufacturing system adopts a solid-liquid interface additive manufacturing layout with steel in the front and aluminum in the back; the method includes the following steps:

[0008] The current process parameters are obtained and additive manufacturing is performed based on the current process parameters. The process parameters include welding torch spacing, aluminum welding torch current and aluminum welding torch voltage. Interface temperature field data is obtained during the manufacturing process, and temperature characteristic parameters are determined based on the interface temperature field data.

[0009] The temperature characteristic parameters are input into the prediction model to obtain the interface brittle phase characteristics; based on the interface brittle phase characteristics, the interface bonding strength is determined.

[0010] In response to the interface bonding strength not meeting the preset threshold, the reverse adjustment model is invoked to determine the adjustment amount of the process parameters;

[0011] Adjust the additive manufacturing process parameters based on the aforementioned process parameter adjustment amount.

[0012] The above solution constructs a closed-loop control system that includes a predictive model and a reverse adjustment model. It uses real-time acquired temperature characteristic parameters to predict the characteristics of the brittle phase at the interface and the bonding strength. When the strength does not meet the requirements, it automatically calculates and adjusts the process parameters, thus solving the problem that traditional open-loop control cannot respond to changes in interface state in real time and realizing proactive control of interface quality.

[0013] Furthermore, the temperature characteristic parameters include the average cooling rate and the high-temperature residence time.

[0014] The above scheme accurately characterizes the interfacial thermal history by selecting the average cooling rate and high-temperature residence time as key feature parameters, providing a reliable data foundation for subsequent accurate prediction of the interfacial microstructure evolution.

[0015] Furthermore, the prediction model includes a first neural network and a second neural network;

[0016] The temperature characteristic parameters are input into the prediction model to obtain the brittle phase characteristics of the interface, including:

[0017] The average cooling rate is input into the first neural network, which is used to predict the composition type of the brittle phase at the interface based on the average cooling rate.

[0018] The high-temperature residence time is input into the second neural network, which is used to predict the layer thickness of the brittle phase at the interface based on the high-temperature residence time.

[0019] The above scheme achieves decoupled prediction of brittle phase composition type and layer thickness by training the first neural network and the second neural network separately, and improves prediction accuracy by utilizing the powerful nonlinear mapping capability of neural networks.

[0020] Furthermore, the prediction model also includes a third neural network;

[0021] Based on the aforementioned brittle phase characteristics of the interface, the interfacial bonding strength is determined, including:

[0022] The composition type and layer thickness of the brittle phase at the interface are input into the third neural network, which is used to predict the interfacial bonding strength based on the composition type and layer thickness of the brittle phase at the interface.

[0023] The above scheme establishes a mapping relationship between microscopic organizational characteristics (composition type, layer thickness) and macroscopic mechanical properties (bonding strength) through a third neural network, realizing accurate deduction from microscopic characteristics to macroscopic properties.

[0024] Furthermore, the inputs to the reverse adjustment model include the current process parameters, temperature characteristic parameters, and interface bonding strength, and the output is the adjustment amount of the process parameters.

[0025] The above solution achieves rapid and accurate inversion of process parameters by constructing a reverse adjustment model with the current state as input and the adjustment amount as output, thus avoiding the blindness of traditional trial-and-error adjustment methods.

[0026] Furthermore, based on the aforementioned process parameter adjustment amount, the additive manufacturing process parameters are adjusted, including:

[0027] The optimized process parameters are calculated based on the adjustment amount of the process parameters. Boundary verification is performed on the optimized parameters to ensure that the parameters are within the preset feasible range. The parameter adjustment is then performed through the dual-gun electric arc additive manufacturing system.

[0028] The above solution ensures that the adjusted process parameters are always within a feasible range for stable forming by introducing a boundary verification mechanism, thus guaranteeing the stability of the additive manufacturing process.

[0029] Furthermore, the first neural network is a 3-layer backpropagation neural network, specifically including a 16-dimensional first hidden layer, an 8-dimensional second hidden layer, and a 4-dimensional third hidden layer;

[0030] The second neural network is a 3-layer backpropagation neural network, specifically including a 16-dimensional fourth hidden layer, an 8-dimensional fifth hidden layer, and a 4-dimensional sixth hidden layer.

[0031] The above scheme balances the prediction accuracy and computational efficiency of the model by limiting the specific hierarchical structure of the first and second neural networks, making it suitable for the real-time inference needs of industrial sites.

[0032] Furthermore, the third neural network is a 3-layer backpropagation neural network, specifically including a 16-dimensional seventh hidden layer, an 8-dimensional eighth hidden layer, and a 4-dimensional ninth hidden layer.

[0033] The above scheme ensures the accuracy of mapping from microscopic features to macroscopic intensity by limiting the specific structure of the third neural network.

[0034] Furthermore, the backpropagation model is a 3-layer backpropagation neural network, specifically consisting of a 16-dimensional tenth hidden layer, an 8-dimensional eleventh hidden layer, and a 4-dimensional twelfth hidden layer.

[0035] The above scheme enhances the model's ability to handle complex nonlinear inverse problems and improves the accuracy of process parameter adjustments by designing a deeper inverse adjustment network.

[0036] This invention also discloses a steel-aluminum dissimilar metal additive manufacturing interface control device for implementing the above method, comprising:

[0037] The data acquisition unit is used to acquire the current process parameters and perform additive manufacturing based on the current process parameters. During the manufacturing process, it acquires interface temperature field data and determines temperature characteristic parameters based on the interface temperature field data.

[0038] The feature parameter prediction unit is used to input the temperature feature parameters into the prediction model to obtain the interface brittle phase characteristics; and to determine the interface bonding strength based on the interface brittle phase characteristics.

[0039] The process parameter adjustment unit is used to respond to the interface bonding strength not meeting the preset threshold by calling the reverse adjustment model, determining the process parameter adjustment amount, and adjusting the additive manufacturing process parameters based on the process parameter adjustment amount.

[0040] The aforementioned device, through the coordinated operation of its various units, realizes the hardware implementation of the interface control method for steel-aluminum additive manufacturing, ensuring the real-time performance and reliability of the method.

[0041] Compared with the prior art, the present invention has the following advantages:

[0042] This invention acquires interface temperature field data in real time and extracts temperature characteristic parameters. It then uses a predictive model to accurately infer the characteristics and bonding strength of the brittle phase at the interface. Furthermore, it calculates the adjustment amount of process parameters by adjusting the model in reverse, forming a complete closed-loop control system. This method can respond to changes in interface state in real time, actively adjust process parameters, and effectively induce the brittle phase to exhibit ideal morphology and thickness. It avoids the problems of parameter adjustment lag and blindness in traditional open-loop control, significantly improving the interface bonding strength and forming quality of steel-aluminum dissimilar metal additive manufacturing components. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is a flowchart of a method for controlling the interface in steel-aluminum dissimilar metal additive manufacturing, as described in an embodiment of the present invention.

[0045] Figure 2 This is a schematic diagram of the dual-gun electric arc additive manufacturing system in an embodiment of the present invention.

[0046] Figure 3 This is a schematic diagram of the gimbal device structure of the servo imaging instrument in an embodiment of the present invention.

[0047] Figure 4 This is a schematic diagram of the three-level cascaded neural network structure of the first neural network in an embodiment of the present invention;

[0048] Figure 5 This is a schematic diagram of the three-level cascaded neural network structure of the second neural network in an embodiment of the present invention;

[0049] Figure 6 This is a schematic diagram of the three-level cascaded neural network structure of the third neural network in an embodiment of the present invention;

[0050] Figure 7 This is a schematic diagram of the three-stage cascaded structure of the reverse regulation model neural network in an embodiment of the present invention;

[0051] Figure 8 This is a schematic diagram of data interaction between neural networks in an embodiment of the present invention.

[0052] The components include: 1. Robot; 2. Dual-gun electric arc additive manufacturing device; 3. Front-mounted steel electric arc gun; 4. Rear-mounted aluminum electric arc gun; 5. Follow-up gimbal; 6. Thermal imager; 7. Component to be formed; 8. Worktable. Detailed Implementation

[0053] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0054] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0055] Example 1

[0056] This embodiment provides a method for controlling the interface in steel-aluminum dissimilar metal additive manufacturing, which is applied to a dual-gun arc additive manufacturing system. The system employs a solid-liquid interface additive manufacturing layout with steel in front and aluminum behind, meaning the steel welding gun deposits the steel layer first, and the aluminum welding gun deposits the aluminum layer later. During aluminum deposition, the steel layer remains solid, thus forming a solid-liquid reaction interface.

[0057] The steel-aluminum front and aluminum rear arrangement is the foundation for interface control in this embodiment. In the additive manufacturing process of dissimilar metals like steel and aluminum, if two liquid molten pools are superimposed, the significant differences in thermophysical properties between steel and aluminum can easily trigger violent interfacial reactions, generating a large amount of brittle intermetallic compounds. This embodiment controls the timing and spatial position so that the first deposited steel layer has cooled to a solid state by the time the aluminum welding torch arrives, with only the aluminum wire melting to form a liquid molten pool. This solid-liquid interface mode effectively suppresses the intensity of the interfacial reaction, providing a stable physical prerequisite for subsequent control of the interface microstructure through process parameters. Figure 2 The hardware configuration of the system is shown, such as... Figure 2 and Figure 3 As shown, the dual-gun arc additive manufacturing system specifically includes a robot 1, a dual-gun arc additive manufacturing device 2, a follow-up gimbal 5, and a thermal imager 6. The robot 1 is positioned on one side of the worktable 8. The dual-gun arc additive manufacturing device 2 is installed at the end effector position of the robot 1. The dual-gun arc additive manufacturing device 2 includes a front steel arc gun 3 and a rear aluminum arc gun 4, arranged sequentially along the additive manufacturing direction to form a solid-liquid interface additive manufacturing layout with steel in front and aluminum behind. The follow-up gimbal 5 is connected to the dual-gun arc additive manufacturing device 2, and the thermal imager 6 is mounted on the follow-up gimbal 5. The follow-up gimbal 5 drives the thermal imager 6 to move synchronously with the dual-gun arc additive manufacturing device 2, collecting temperature field data of the interface area in real time. The component to be formed 7 is placed on the worktable 8, and the robot 1 drives the dual-gun arc additive manufacturing device 2 to perform additive manufacturing on the component to be formed 7 along a preset path.

[0058] The method in this embodiment specifically includes the following steps:

[0059] Step S100: Obtain the current process parameters and perform additive manufacturing based on the current process parameters. The process parameters include welding torch spacing, aluminum welding torch current and aluminum welding torch voltage. During the manufacturing process, obtain interface temperature field data and determine temperature characteristic parameters based on the interface temperature field data.

[0060] Process parameters are key variables determining interfacial heat input and thermal cycling. The welding torch spacing determines the spatial distance between the steel layer and the molten aluminum pool, thus affecting the interpass temperature; the welding torch current and voltage directly determine the arc power and heat input. During additive manufacturing, a follow-up thermal imager acquires infrared thermal images of the interface region in real time. After filtering, noise reduction, and emissivity correction, a continuous temperature change curve is extracted. Based on this curve, the system further calculates temperature characteristic parameters that characterize the interface thermal history. It should be understood that although this embodiment preferably uses welding torch spacing, current, and voltage as core control parameters, in other embodiments, parameters such as wire feed speed and travel speed can also be introduced as auxiliary inputs, as long as these parameters can have a substantial impact on the interface temperature field.

[0061] Specifically, the temperature characteristic parameters include the average cooling rate and the high-temperature residence time. These two parameters are selected as the core inputs to the prediction model based on the growth mechanism of intermetallic compounds at the steel-aluminum interface. The average cooling rate directly determines the undercooling and solidification rate of the interfacial melt, thus affecting the nucleation and growth kinetics of the brittle phase. For example, a higher cooling rate can suppress long-range atomic diffusion, helping to inhibit the excessive growth of tongue-shaped or needle-shaped Fe2Al5 phases, thereby inducing the formation of spherical FeAl3 phases with less negative impact on interfacial strength. The high-temperature residence time characterizes the time span during which the interface remains in a high-temperature active state, directly related to the interdiffusion flux of Fe and Al atoms. An excessively long residence time leads to sufficient atomic diffusion, increasing the cumulative thickness of the brittle phase layer and easily triggering interfacial delamination; an excessively short residence time may result in insufficient diffusion, failing to form a continuous metallurgical bonding layer. Therefore, by acquiring these two characteristic parameters in real time, the thermal history of the interfacial microstructure evolution can be accurately captured.

[0062] Step S200: Input the temperature characteristic parameters into the prediction model to obtain the brittle phase characteristics of the interface.

[0063] The predictive model is the core intelligent component of this embodiment, establishing a mapping relationship from macroscopic temperature characteristics to microscopic structural characteristics. The characteristics of the brittle phase at the interface mainly include the composition type of the brittle phase (such as FeAl3 or Fe2Al5) and the thickness of the brittle phase layer. By learning from a large amount of experimental data, the predictive model can quickly infer the type and thickness of the brittle phase that may form at the interface under the current process conditions based on the input temperature characteristic parameters. This step transforms microscopic structural characteristics that are difficult to measure in real time into predictable values ​​that can be calculated in real time, solving the problem of lag in traditional methods that rely on post-process metallographic inspection.

[0064] Specifically, the prediction model includes a first neural network and a second neural network. For example... Figure 4 and Figure 5 As shown, this embodiment adopts a decoupled prediction strategy, which decomposes the complex temperature-tissue mapping process into two parallel subtasks, each undertaken by two independent neural networks.

[0065] The average cooling rate is input into a first neural network, which predicts the composition type of the brittle phase at the interface based on the average cooling rate. The first neural network is trained on a large number of samples to establish a nonlinear mapping relationship between cooling rate and phase composition type. Specifically, the output layer of the first neural network outputs the probability values ​​of different types of brittle phases at the interface, such as the probability of "FeAl3 predominant" and the probability of "Fe2Al5 predominant". The system selects the type with the highest probability as the current composition type prediction result. This classification prediction method allows the system to determine in real time whether the current process conditions are conducive to generating the ideal brittle phase morphology. Specifically, the first neural network is a three-layer backpropagation neural network, with a specific structure including a 16-dimensional first hidden layer, an 8-dimensional second hidden layer, and a 4-dimensional third hidden layer. Figure 4 As shown, the 1D feature (average cooling rate) of the input layer first enters the 16D first hidden layer. This layer has a large number of neurons, designed to provide sufficient feature space to capture the complex nonlinear relationship between cooling rate and brittle phase kernel. Subsequently, the data flows through the 8D second hidden layer and the 4D third hidden layer, extracting higher-order abstract features layer by layer, and finally outputting the probability distribution of brittle phase composition types in the output layer. It should be understood that although this embodiment preferably uses a 3-layer hidden layer structure, in other embodiments, 2 or 4 layers of hidden layers can also be used, as long as an effective mapping from cooling rate to phase type can be achieved. However, after extensive experimental verification, the 3-layer structure effectively controls the number of network parameters while ensuring mapping accuracy, avoiding the gradient vanishing problem caused by excessive network depth.

[0066] Simultaneously, the high-temperature residence time is input into a second neural network, which predicts the layer thickness of the brittle phase at the interface based on the high-temperature residence time. The second neural network performs a regression task, and its output is a continuous numerical value representing the predicted brittle phase layer thickness. Specifically, the second neural network is a three-layer backpropagation neural network, with a structure including a 16-dimensional fourth hidden layer, an 8-dimensional fifth hidden layer, and a 4-dimensional sixth hidden layer. Figure 5 As shown, the structure of the second neural network is consistent with that of the first neural network, both employing a 16-8-4 hidden layer dimension configuration. This consistent design is not accidental, but rather based on the consideration that the mapping relationship between high-temperature dwell time and layer thickness also exhibits highly nonlinear characteristics. The 16-dimensional input layer can fully receive and process the time-series features of the dwell time, while the subsequent compression layer is responsible for extracting the key dynamic parameters that determine layer thickness growth. This embodiment avoids blind spots in network structure design by limiting specific dimensions, ensuring that the model has stable feature extraction capabilities when processing different input features.

[0067] Step S300: Determine the interfacial bonding strength based on the characteristics of the brittle phase at the interface.

[0068] Interfacial bonding strength is a core indicator for evaluating the quality of steel-aluminum composite components. This embodiment establishes a correlation model between brittle phase characteristics and bonding strength, enabling the deduction from microstructure to macroscopic properties. Different brittle phase types and layer thicknesses have different mechanisms of influence on strength: for example, the Fe2Al5 phase is needle-like or tongue-like, with high hardness but poor toughness, easily causing stress concentration and significantly reducing bonding strength; while the FeAl3 phase tends to form spherical shapes, having a smaller negative impact on the interface. By predicting the brittle phase characteristics output by the model, the system can assess in real time whether the current interfacial bonding strength meets the design requirements.

[0069] The predictive model also includes a third neural network. For example... Figure 6 As shown, after obtaining the two microscopic features of the composition type and layer thickness of the brittle phase at the interface, the system uses these two features as inputs and passes them to the third neural network.

[0070] Specifically, the composition type and layer thickness of the brittle phase at the interface are input into a third neural network. This third neural network predicts the interfacial bonding strength based on these factors. The third neural network constructs a structure-performance mapping model from microscopic features to macroscopic mechanical properties. Since interfacial bonding strength is not a linear function of a single factor, but rather the result of the coupling effect between the brittle phase type (morphology, hardness) and layer thickness—for example, with the same layer thickness of 10 μm, the relative strength reduction of FeAl3 and Fe2Al5 phases is drastically different—the third neural network, by learning from a large amount of experimental data, can accurately capture this complex coupling relationship and output highly accurate predicted values ​​for interfacial bonding strength. This three-level cascaded prediction architecture, where the first and second networks predict microscopic features separately, and the third network comprehensively predicts macroscopic properties, effectively avoids the potential loss of physical meaning and model overfitting problems that may occur when directly predicting strength from temperature parameters, significantly improving the accuracy and interpretability of the prediction results.

[0071] Furthermore, the third neural network is a 3-layer backpropagation neural network, specifically consisting of a 16-dimensional seventh hidden layer, an 8-dimensional eighth hidden layer, and a 4-dimensional ninth hidden layer. For example... Figure 6 As shown, the third neural network is responsible for predicting macroscopic mechanical properties by fusing microscopic organizational features. Its input layer is 2-dimensional (composition type and layer thickness), and the hidden layers also follow the 16-8-4 dimensionality reduction logic. The core advantage of this structural design lies in balancing the risk of overfitting with expressive power. If the hidden layer dimension is too high (such as 64 or 128 dimensions), although the model's fitting ability is enhanced, it is very easy to learn noise in the training data, leading to a decrease in generalization ability under unknown working conditions; if the dimension is too low (such as 4 or 2 dimensions), it is difficult to capture the coupling effect mechanism of composition type and layer thickness on strength. The 16-8-4 structure selected in this embodiment, through the layer-by-layer decreasing dimension design, forces the network to learn the most representative feature combination, thereby achieving high-precision strength prediction with limited training samples.

[0072] The reason for choosing the specific layer and dimension configuration mentioned above in this embodiment is based on the following three technical considerations:

[0073] First, overfitting is avoided. Although the training samples for steel-aluminum additive manufacturing have been expanded, they are still small compared to the million-level datasets commonly used in deep learning. Employing a three-layer hidden structure with moderate dimensionality effectively limits the order of magnitude of model parameters. Combined with early stopping mechanisms and regularization strategies, this significantly reduces the risk of a model performing well on the training set but failing in real-world applications.

[0074] Secondly, it ensures nonlinear mapping capability. The generation and evolution of brittle phases at the interface involve multi-physics coupling processes such as thermodynamics and kinetics, which cannot be accurately described by simple linear models. The three-layer hidden layer structure in this embodiment, combined with nonlinear activation functions such as ReLU, endows the network with powerful nonlinear fitting capabilities, which can approximate arbitrarily complex continuous function relationships, thereby accurately characterizing the mapping relationship between temperature features and microstructure, and between microstructure and macroscopic properties.

[0075] Third, it meets the requirements for real-time inference speed. In the process of arc additive manufacturing, the adjustment of process parameters must be real-time or near real-time; otherwise, the meaning of closed-loop control will be lost. Compared with deep networks (such as ResNet, VGG, etc.), the shallow wide network structure used in this embodiment has a smaller computational load and lower inference latency. Actual measurements show that the inference time of a single neural network can be controlled at the millisecond level, fully meeting the timeliness requirements of online monitoring and real-time control in additive manufacturing, and providing a computational foundation for the rapid response of the subsequent reverse adjustment model.

[0076] Step S400: In response to the interface bonding strength not meeting the preset threshold, the reverse adjustment model is invoked to determine the adjustment amount of the process parameters; based on the adjustment amount of the process parameters, the additive manufacturing process parameters are adjusted.

[0077] This step establishes a complete closed-loop control logic. When the interface bonding strength determined in step S300 is lower than a preset safety threshold (e.g., 180 MPa), the system determines that the current process state cannot meet the quality requirements and triggers the reverse adjustment model. Based on the current interface state and the target strength, this model reversely derives the adjustment direction and magnitude of the process parameters, outputting specific adjustment amounts (e.g., current increase, welding torch spacing adjustment). The system executes this adjustment command, correcting the additive manufacturing process in real time, thereby optimizing the interface structure in subsequent deposition layers. This perception-prediction-adjustment closed-loop mechanism overcomes the shortcomings of traditional open-loop control in handling operating condition fluctuations (e.g., changes in ambient temperature, changes in heat dissipation conditions), achieving proactive control of interface quality.

[0078] When the predicted interfacial bonding strength is lower than the preset safety threshold, the system needs to quickly and accurately calculate the adjustment amount of the process parameters in order to correct the subsequent additive manufacturing process.

[0079] Specifically, this embodiment uses a reverse adjustment model to determine the adjustment amount of the process parameters. The inputs to the reverse adjustment model include the current process parameters, temperature characteristic parameters, and interface bonding strength, and the output is the adjustment amount of the process parameters. It should be understood that deriving process parameters from interface performance is a typical many-to-many mapping problem with highly nonlinear characteristics. For example, the same insufficient strength may be caused by excessively rapid cooling, insufficient residence time, or a combination of both. By introducing the current process parameters (welding torch spacing, current, voltage) and temperature characteristic parameters (cooling rate, residence time) as joint inputs, the reverse adjustment model can comprehensively determine the root cause of insufficient strength, thereby outputting a precise adjustment amount. This data-driven reverse mapping mechanism effectively solves the problems of traditional trial-and-error methods relying on human experience, lagging adjustments, and susceptibility to oscillations, achieving intelligent closed-loop optimization of process parameters.

[0080] After obtaining the process parameter adjustment amount, the system does not execute it directly, but instead executes a rigorous set of verification and execution logic. Based on the process parameter adjustment amount, the additive manufacturing process parameters are adjusted, including: calculating the optimized process parameters according to the process parameter adjustment amount, performing boundary verification on the optimized parameters to ensure that the parameters are within the preset feasible range, and executing the parameter adjustment through the dual-gun electric arc additive manufacturing system.

[0081] Specifically, such as Figure 8 As shown, the optimized process parameters are obtained by superimposing the adjustment values ​​(such as Δd, ΔI, ΔU) output by the reverse adjustment model with the current parameters. However, the values ​​output by the neural network may theoretically exceed the physical equipment's tolerance limits or the process stability window. For example, the calculated aluminum welding torch current may exceed the maximum output current of the power supply, or the welding torch spacing may be too small, causing interference from the two torches. Therefore, this embodiment introduces a crucial boundary verification step. The system presets feasible ranges for each parameter, such as a feasible range of 5~20mm for the welding torch spacing, 150~200A for the aluminum welding torch current, and 18~24V for the aluminum welding torch voltage. Before performing the adjustment, the system determines whether the optimized parameters are within these ranges. If the parameters overflow, the system will automatically clamp them at the boundary values ​​or trigger an alarm to prompt manual intervention, thereby preventing process failures or even equipment damage caused by parameter overflow. This boundary verification mechanism reflects the engineering practicality considerations of this invention, ensuring the safety and robustness of the intelligent control algorithm in industrial field applications. Finally, the verified safety parameters will be sent to the control unit of the dual-gun arc additive manufacturing system via the communication interface to adjust the welding gun posture and power output in real time, thus completing closed-loop control.

[0082] In this embodiment, the first-third neural network and the inverse adjustment model are trained through the following operations.

[0083] First, a dataset was constructed and preprocessed. A dedicated training dataset for steel-aluminum additive manufacturing was built, containing 100,000 standardized experimental samples. This dataset covers the entire process range, including cooling rates of 50-250℃ / s and high-temperature dwell times above 250℃ of 2-20s, as well as labeled data on the corresponding brittle phase composition types, phase layer thicknesses, and interface bonding strengths. This fully covers all working conditions, including different combinations of steel and aluminum materials, different environmental heat dissipation conditions, and different component structures. The dataset underwent standardization preprocessing: First, the input features were normalized by mapping the cooling rate and high-temperature dwell time to the [0,1] interval to eliminate the influence of dimensional differences on model training. The normalization formula is as follows:

[0084]

[0085] in, , These are the global minimum and maximum values ​​for the corresponding features, obtained through extensive experimental statistics in the early stages.

[0086] The dataset was then randomly divided into training, validation, and test sets in an 8:1:1 ratio: the training set was used for iterative updates of model weights; the validation set was used for hyperparameter tuning and early stopping detection during training; and the test set consisted of independent samples not involved in training, used for final model accuracy verification.

[0087] Next, the first neural network is pre-trained. The hidden layers of the first neural network use the ReLU activation function to solve the gradient vanishing problem. The formula is:

[0088]

[0089] The output layer uses the Softmax activation function to transform the output into probabilities for each category, ensuring that the sum of the probabilities is 1. The formula is as follows:

[0090]

[0091] The network weights are initialized using the Xavier uniform initialization method to ensure that the variance of the input and output is consistent, avoiding the gradient vanishing or exploding problems in the early stages of training. The weight initialization range is:

[0092]

[0093] in, For the corresponding layer's input dimension, The bias term is initialized to 0 for the output dimension of the corresponding layer.

[0094] The training process employs the Adam adaptive learning rate optimizer to achieve adaptive weight updates. Its parameter update formula is as follows:

[0095]

[0096]

[0097]

[0098] in, This represents the loss gradient for the current round. , The attenuation coefficients for the first and second moments are denoted as . To ensure numerical stability, avoid denominators of 0; initial learning rate. The learning rate decreases to 0.9 times its original value after every 50 training rounds, gradually converging to the optimal solution. The loss function used is cross-entropy loss, adapted to the optimization objective of the classification task, and the formula is:

[0099]

[0100] in, For the true label of the sample, This represents the predicted probability of the model. An early stopping mechanism is introduced during training: training is automatically terminated when the loss on the validation set fails to decrease for 10 consecutive rounds to prevent overfitting. After final pre-training, the first neural network achieves a phase composition classification accuracy of 98.5%, accurately predicting the composition type of brittle phases at interfaces.

[0101] Then, the second neural network is pre-trained. The hidden layers of the second neural network use the ReLU activation function, and the output layer uses a linear activation function to adapt to the continuous output of the regression task. Weight initialization also uses Xavier uniform initialization, and the optimizer is the Adam adaptive learning rate optimizer. The parameter update rule is the same as that of the first network, and the initial learning rate is set to... A learning rate decay mechanism is included. The loss function uses mean squared error (MSE) loss, adapted to the optimization objective of the regression task, and the formula is:

[0102]

[0103] in, The true layer thickness of the sample. This represents the predicted layer thickness of the model. An early stopping mechanism is also introduced, terminating training when the validation set loss fails to decrease for 10 consecutive rounds. After pre-training, the mean absolute error (MAE) and root mean square error (RMSE) of the second neural network's layer thickness prediction are ≤0.5μm and ≤0.7μm, respectively, meeting the process requirements for prediction accuracy.

[0104] Subsequently, a third neural network was pre-trained. The hidden layers of this third neural network used the ReLU activation function, and the output layer used a linear activation function. Weights were initialized using Xavier uniform initialization, and the optimizer was the Adam adaptive learning rate optimizer. The initial learning rate was... The learning rate decays accordingly. The loss function uses mean squared error loss, and an early stopping mechanism is also introduced during training to avoid overfitting. After pre-training, the mean absolute error (MAE) of the third neural network's interface strength prediction is ≤5MPa, and the coefficient of determination R² ≥0.97, accurately mapping microstructural features to macroscopic mechanical properties.

[0105] After pre-training the three neural networks, joint fine-tuning and model accuracy verification were performed. The three pre-trained neural networks were concatenated into a three-level concatenated intelligent prediction model. The overall input was cooling rate and high-temperature dwell time, and the overall output was interface bonding strength. First, the weights of the first two hidden layers of the first two networks were frozen, and only the weights of the later layers were fine-tuned for initial adaptation to reduce the bias of pre-trained features. Then, the weights of all networks were unfrozen, and joint fine-tuning was performed using the full training set samples to further optimize the weight parameters of the overall model, reduce cascade errors, and improve overall prediction accuracy.

[0106] Finally, the parameter adjustment inverse model is trained and the adjustment amount is calculated. In order to realize a complete prediction-adjustment closed-loop mechanism, an additional parameter adjustment inverse model is trained. This model is used to accurately calculate the adjustment amount of process parameters based on the current interface state and target performance, which solves the problem of blind parameter adjustment in traditional open-loop control. The model is built based on the mapping relationship of the forward three-level prediction model. Using the prior knowledge of the forward model, a reverse performance-parameter mapping relationship is constructed. The specific process is as follows: Traversing the feasible range of process parameters: The feasible range of welding torch spacing is set to 5~20mm, the feasible range of aluminum welding torch current is set to 150~200A, and the feasible range of aluminum welding torch voltage is set to 18~24V, covering all feasible process ranges in industrial settings to ensure that the adjusted parameters remain within a stable range. Generating forward mapping samples: For all feasible combinations of process parameters, the forward three-level prediction model is called to calculate the corresponding cooling rate, high-temperature residence time, and interface bonding strength, generating a total of 100,000 complete forward mapping samples. Each sample includes: process parameters (welding torch spacing d, aluminum welding torch current I, aluminum welding torch voltage U) → interface state (cooling rate). High temperature dwell time → Interface strength σ;

[0107] Based on the positive samples, we construct the negative training samples, with the input of the samples being: the current cooling rate. Current high temperature stay time Current process parameters ( , , ), target interface strength ;

[0108] The sample output is: the adjustment amount of the process parameters (Δd, ΔI, ΔU), where Δd = ΔI= ΔU= This refers to the difference between the adjusted parameters and the current parameters, ensuring that the model directly outputs the executable adjustment amount without additional calculation.

[0109] The parameter adjustment inverse model is a 3-layer backpropagation neural network (3 hidden layers, with 16-dimensional, 8-dimensional, and 4-dimensional neurons respectively). Its network structure is as follows: the input layer consists of 3-dimensional neurons, corresponding to the normalized current cooling rate, the current high-temperature dwell time, and the target interface intensity; the hidden layers consist of 16-dimensional, 8-dimensional, and 4-dimensional neurons respectively, and the hidden layers use the ReLU activation function, with the formula:

[0110]

[0111] It is used to fit complex nonlinear inverse mapping relationships; the output layer has 3D neurons, which correspond to the welding torch spacing adjustment Δd, the aluminum welding torch current adjustment ΔI, and the aluminum welding torch voltage adjustment ΔU, respectively. The output layer uses a linear activation function to adapt to continuous adjustment output.

[0112] The network weights are also initialized uniformly using Xavier, and the training process uses the Adam adaptive learning rate optimizer. The parameter update rule is the same as that of the forward model, and the initial learning rate is... A learning rate decay mechanism is included. The loss function uses mean squared error loss, and the errors of the three adjustment parameters are optimized. The formula is as follows:

[0113]

[0114] The training process also incorporates an early stopping mechanism, terminating training when the validation set loss fails to decrease for 10 consecutive rounds to avoid overfitting.

[0115] After training, the model can complete the calculation of the adjustment amount within 10ms. The specific execution logic is as follows: (1) Set the current interface state and process parameters ( , , (1) Input the target intensity into the adjustment model; (2) The model outputs three adjustment values: Δd, ΔI, and ΔU, which are the adjustment differences of the three parameters respectively; (3) Calculate the adjusted process parameters; (4) Perform boundary verification on the adjusted parameters to ensure that the parameters are within the process feasible range.

[0116] Optionally, through the aforementioned regulation, the thickness of the brittle phase at the interface can be controlled within the optimal range of 5~8μm, and the brittle phase can be induced to transform from the conventional tongue-shaped and needle-shaped FeAl3 phase into a spherical phase, effectively reducing the stress concentration at the interface and increasing the interfacial bonding strength by more than 35%.

[0117] Optionally, the dataset used in this invention specifically comprises: constructing 100,000 sets of standardized experimental samples for steel-aluminum additive manufacturing, covering cooling rates of 50~250℃ / s and high-temperature dwell times of 2~20s. Each set includes input features such as cooling rate and high-temperature dwell time, as well as labeled data on brittle phase composition, layer thickness, and interface strength. After normalization, the samples are divided into training, validation, and test sets in a ratio of 8:1:1. Through 300~500 sets of basic experiments (to determine parameter ranges and calibrate true values) and 500~1000 sets of supplementary experiments (to expand the sample), the 100,000 sets of samples can be expanded and generated based on the basic experimental data.

[0118] The following specific application examples will further illustrate the scheme and effects of the method of the present invention.

[0119] The electric arc additive manufacturing experiment in this implementation scheme uses the following basic parameters to provide unified data support for the training, inference, and verification of the four neural networks, such as... Figure 1 The following is a detailed flowchart of the implementation scheme of this patent.

[0120] In step S01, first, prepare an arc additive manufacturing steel substrate, made of Q235 low carbon steel, 10mm thick, with the surface polished to remove oxide scale, and dimensions of 300mm × 100mm × 10mm; a twin-wire welding torch device (such as...) Figure 2-3 (As shown) The front steel welding torch uses ER50-6 wire, and the rear aluminum welding torch uses 5356 aluminum alloy wire. Arc additive manufacturing process parameters: arc current 180A, arc voltage 22V, welding torch movement speed 6mm / s (adjustable range 5~8mm / s), wire feed speed 8m / min, shielding gas Ar, flow rate 20L / min. Cooling rate measurement: An infrared thermal imager (measurement accuracy ±2℃ / s) was used to collect the average cooling rate of the interface region from 250℃ to room temperature, with a measurement range of 50~250℃ / s. High-temperature dwell time measurement: The high-temperature dwell time above 250℃ in the interface region was simultaneously collected, with a measurement range of 2~20s. Scanning electron microscopy (SEM, magnification 5000×) was used to observe the composition and morphology of the brittle phases at the interface. Combined with energy dispersive spectroscopy (EDS), the formation and layer thickness of FeAl3 and Fe2Al5 were determined, verifying the neural network prediction results. The interfacial bond strength was tested using an electronic universal testing machine (range 0~500kN, accuracy ±0.1MPa) through shear tensile tests.

[0121] In steps S02 and S03, the cooling rate is first normalized to the interval [0,1]. =50℃ / s =250℃ / s; normalized interval for high-temperature residence time [0,1]. =2s、 =20s; normalized interval for interface bonding strength [0,1], =50MPa =250MPa; welding torch spacing normalized interval [0,1] =5mm =20mm; Normalized range of aluminum welding torch current [0,1] =150A、 =200A; Normalized voltage range for aluminum welding torch [0,1] =18V =24V.

[0122] One set of stacking parameters was selected and used throughout the four neural network inference processes as unified example data: actual average cooling rate 90℃ / s, actual high temperature dwell time 8s, current welding torch spacing 12mm, current aluminum welding torch current 180A, current aluminum welding torch voltage 22V, and target interface strength 180MPa to demonstrate and analyze the neural network derivation process.

[0123] like Figure 4 As shown, in S02, the first neural network is a 3-layer backpropagation (BP) classification network. Its core function is to predict the composition type of the brittle phase of the interface based on the average cooling rate. The structure is 1-dimensional input → 16-dimensional hidden layer 1 → 8-dimensional hidden layer 2 → 4-dimensional hidden layer 3 → 2-dimensional output. The matrices of each layer and the input and output results are as follows:

[0124] The input is the average cooling rate (1D), which is normalized before being input into the network. The preprocessing formula is:

[0125]

[0126] The first hidden layer of the first neural network (16 dimensions)

[0127] Substituting the example data, the normalized input value is 0.16, and the input vector (1×1) is: h0 = [0.16].

[0128] Weight matrix W1 (16×1, Xavier uniform initialization, range) =[-0.594,0.594].

[0129] The specific values ​​are as follows:

[0130]

[0131] Bias term b1 (16×1, initialized to 0):

[0132]

[0133] Activation function: After activation, the output vector is h1 (16×1):

[0134]

[0135] The second hidden layer of the first neural network (8 dimensions)

[0136] The weight matrix W2 (8×16, Xavier uniform initialization) has the following values:

[0137]

[0138] Bias term b2 (8×1, initialized to 0):

[0139]

[0140] Activation function: ReLU, output vector h2 (8×1, rounded to 2 decimal places) after activation:

[0141]

[0142] The third hidden layer of the first neural network (4-dimensional)

[0143] The weight matrix W3 (4×8, uniformly initialized by Xavier) has the following values:

[0144]

[0145] Bias term b3 (4×1, initialized to 0):

[0146]

[0147] Activation function: ReLU, output vector h3 (4×1, rounded to 2 decimal places) after activation:

[0148]

[0149] First neural network output layer: Prediction of brittle phase composition

[0150] The weight matrix W4 (2×4, uniformly initialized by Xavier) has the following values:

[0151]

[0152] Bias term b4 (2×1, initialized to 0):

[0153]

[0154] Activation function: Softmax, output vector h4 (2×1, rounded to 2 decimal places) after activation:

[0155]

[0156] Output meaning: h4(1)=0.97 (probability of FeAl3 being the main component), h4(2)=0.03 (probability of Fe2Al5 being the main component), the prediction result is "FeAl3 being the main component", which is consistent with the actual detection result.

[0157] like Figure 5 As shown, the second neural network in S03 is a 3-layer BP regression network. Its core function is to predict the thickness of the brittle phase layer at the interface based on the high-temperature residence time. The structure is 1-dimensional input → 16-dimensional hidden layer 1 → 8-dimensional hidden layer 2 → 4-dimensional hidden layer 3 → 1-dimensional output. The matrices of each layer and the input and output results are specific values ​​as follows.

[0158] The input is the high-temperature residence time (1-dimensional), which is normalized before being input into the network. The preprocessing formula is:

[0159]

[0160] Substitute the example data 8s, =2s、 =20s, the normalized input value is calculated to be 0.33, and the input vector (1×1) is: h0=[0.33]

[0161] The first hidden layer of the second neural network (16 dimensions)

[0162] The weight matrix W1 (16×1, Xavier uniform initialization, range [-0.594, 0.594]) has the following specific values:

[0163]

[0164] Bias term b1 (16×1, initialized to 0), activation function: ReLU, output vector h1 (16×1) after activation.

[0165]

[0166] The second hidden layer of the second neural network (8 dimensions)

[0167] The weight matrix W2 (8×16, Xavier uniform initialization) has the following values:

[0168]

[0169] Bias term b2 (8×1, initialized to 0), activation function: ReLU, output vector h2 (8×1) after activation:

[0170]

[0171] The third hidden layer of the second neural network (4-dimensional)

[0172] The weight matrix W3 (4×8, uniformly initialized by Xavier) has the following values:

[0173]

[0174] Bias term b3 (4×1, initialized to 0), activation function: ReLU, output vector h3 (4×1, rounded to 2 decimal places) after activation:

[0175]

[0176] Second neural network output layer

[0177] The weight matrix W4 (1×4, uniformly initialized by Xavier) has the following values:

[0178]

[0179] Bias term b4 (1×1, initialized to 0), activation function: linear activation function, output vector h4 (1×1):

[0180]

[0181] The predicted layer thickness calculated by inverse normalization is 6.1 μm, which deviates from the actual layer thickness of 6.2 μm by 0.1 μm, meeting the accuracy requirement of MAE≤0.5 μm.

[0182] like Figure 6 As shown, the third neural network in S04 is a 3-layer backpropagation (BP) regression network. Its core function is to predict the interface binding strength based on the core features output by the first and second neural networks. The structure is: 3D input → 16D hidden layer 1 → 8D hidden layer 2 → 4D hidden layer 3 → 1D output. The calculation process is similar to the first two neural networks. Through calculation, we can obtain... After inverse normalization, the predicted interfacial bonding strength was calculated to be 178 MPa, which is close to the actual strength of 180 MPa, and meets the accuracy requirements of MAE≤5MPa and R²≥0.97.

[0183] like Figure 7As shown, the parameter adjustment inverse model in S05 is a 3-layer BP neural network. Its core function is to output the adjustment amount (3-dimensional output) of welding torch spacing, aluminum welding torch current, and aluminum welding torch voltage based on the current interface strength, cooling rate, and high temperature dwell time (3-dimensional input). The structure is 3-dimensional input → 16-dimensional hidden layer 2 → 8-dimensional hidden layer 3 → 4-dimensional hidden layer 4 → 3-dimensional output, with each layer matrix and input / output results.

[0184] Calculated After output inverse normalization and adjustment calculation (the adjustment is the change value of the actual parameter corresponding to the normalized output, derived in reverse based on the normalization interval of each parameter):

[0185] Welding torch spacing adjustment: Inverse normalization formula, rounded to 4.7mm (adjustment accuracy 0.1mm);

[0186] Aluminum welding torch current adjustment: Inverse normalization formula, rounded to 16A (adjustment accuracy 1A).

[0187] Aluminum welding torch voltage adjustment: reverse normalization formula, rounded to 1.8V (adjustment accuracy 0.1V).

[0188] Optimized process parameters: Current welding torch spacing 12mm + adjustment amount 4.7mm = 16.7mm; Current aluminum welding torch current 180A + adjustment amount 16A = 196A; Current aluminum welding torch voltage 22V + adjustment amount 1.8V = 23.8V.

[0189] Verification results: Substituting the optimized process parameters into the arc additive manufacturing experiment, the actual interfacial bonding strength was measured to be 182 MPa, which deviated from the target strength of 180 MPa by 2 MPa, meeting the accuracy requirement of MAE≤5 MPa; the brittle phase layer thickness was 5.9 μm, and the brittle phase composition was still mainly FeAl3 (accounting for 91%), and the brittle phase at the interface was effectively controlled, verifying the effectiveness of the parameter adjustment inverse model.

[0190] To ensure the stability and generalization ability of the four neural networks, training and validation details were added, and the reproducibility of the technical solution was further improved. The optimizer used was the Adam optimizer, with a learning rate of 0.001, a decay factor of 0.9, a momentum of 0.99, and a weight decay factor of [missing information]. Loss function: The first neural network uses the cross-entropy loss function, while the second and third neural networks and the parameter adjustment inverse model use the mean squared error (MSE) loss function; Training epochs: 1000 epochs, early stopping strategy (training stops if the validation set loss does not decrease for 50 consecutive epochs), batch size 32; Regularization: L2 regularization is used to prevent overfitting, with a regularization coefficient of 0.0001.

[0191] like Figure 8The diagram shows a specific data transmission and interaction. After the adjustment amount of the process parameters is obtained through the quadruple neural network calculation, the welding torch spacing and arc parameters are adjusted through the coordinated control of the dual-gun device and the arc welding power source, and the next layer of deposition and forming continues until the additive manufacturing of the component is completed.

[0192] Example 2

[0193] This embodiment provides an interface control device for additive manufacturing of dissimilar metals such as steel and aluminum. This device is used to implement the interface control method described in the above embodiment. Figure 2 and Figure 3 As shown, this device can be integrated into an industrial control computer, a programmable logic controller (PLC), or a dedicated embedded controller at the hardware level, serving as the control core of a dual-gun electric arc additive manufacturing system. Logically, the device is mainly divided into a data acquisition unit, a characteristic parameter prediction unit, and a process parameter adjustment unit. These units interact in real-time via a data bus or internal communication protocol, collaboratively achieving closed-loop control of interface quality.

[0194] The data acquisition unit is used to acquire current process parameters and perform additive manufacturing based on these parameters. During manufacturing, it acquires interface temperature field data and determines temperature characteristic parameters based on this data. Specifically, the data acquisition unit is the sensing front end of the entire control system. Its hardware interface is connected to the motion control card and welding power supply of the dual-gun arc additive manufacturing system, real-time reading of process parameters such as the current welding torch spacing, aluminum welding torch current, and aluminum welding torch voltage. Simultaneously, this unit connects to a servo thermal imager via a high-speed data transmission interface (such as Cameralink or GigE interface) to receive infrared thermal image data of the interface area in real time. At the data processing level, the data acquisition unit integrates an image processing algorithm module capable of filtering and denoising the original thermal image, correcting emissivity, and reconstructing the temperature field. Based on the continuous temperature change curve, it calculates the average cooling rate and high-temperature dwell time. It should be understood that although this embodiment preferably uses a thermal imager as the temperature acquisition device, in other embodiments, the data acquisition unit can also connect to multiple thermocouples or pyrometers distributed around the interface to acquire temperature field data through multi-point data fitting, as long as characteristic parameters representing the interface's thermal history can be extracted.

[0195] The feature parameter prediction unit is used to input temperature feature parameters into the prediction model to obtain the characteristics of the brittle phase at the interface; based on the characteristics of the brittle phase at the interface, it determines the interface bonding strength. The feature parameter prediction unit is the intelligent decision-making core of the device, internally storing a trained prediction model, specifically including a first neural network, a second neural network, and a third neural network. This unit receives the average cooling rate and high-temperature residence time output by the data acquisition unit. First, it calls the first neural network to predict the composition type of the brittle phase at the interface, and calls the second neural network to predict the layer thickness of the brittle phase at the interface. Then, it inputs these two microscopic features into the third neural network to comprehensively calculate the current predicted value of the interface bonding strength. In terms of hardware implementation, the feature parameter prediction unit can utilize a CPU for general-purpose calculations, or it can be equipped with a GPU accelerator card or FPGA module to meet the real-time requirements of neural network inference. This unit compares the predicted interface bonding strength value with a preset safety threshold to generate a judgment result signal.

[0196] The process parameter adjustment unit is used to respond to situations where the interface bonding strength does not meet a preset threshold. It invokes a reverse adjustment model to determine the process parameter adjustment amount and adjusts the additive manufacturing process parameters based on this adjustment. The process parameter adjustment unit is the core of the device's execution control. When it receives an insufficient strength signal from the characteristic parameter prediction unit, it immediately activates the reverse adjustment model. This model takes the current process parameters, temperature characteristic parameters, and interface bonding strength as inputs, performs nonlinear mapping calculations, and outputs the welding torch spacing adjustment, aluminum welding torch current adjustment, and aluminum welding torch voltage adjustment. The process parameter adjustment unit also integrates boundary verification logic to determine whether the optimized parameters are within a preset feasible range. If the verification passes, the unit generates corresponding control commands (such as analog voltage signals, digital communication messages, or PWM modulation signals) and sends them to the actuators of the dual-torch arc additive manufacturing system, including a servo motor driver controlling the welding torch spacing and a power controller controlling the welding current and voltage, thereby achieving real-time intervention in the additive manufacturing process.

[0197] Furthermore, the device described in this embodiment also includes a memory for storing data such as the weight parameters, preset thresholds, and feasible ranges of process parameters of the aforementioned prediction model and reverse adjustment model. The processor executes the computer program stored in the memory to control the data acquisition unit, feature parameter prediction unit, and process parameter adjustment unit to work collaboratively, implementing the steps described in Embodiment 1. Through this modular design combining hardware and software, this embodiment solidifies the complex interface intelligent control algorithm into an executable device entity, which not only facilitates deployment and integration in industrial settings but also provides a clear hardware carrier and functional boundaries for subsequent infringement determination.

[0198] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions 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 interface control in steel-aluminum dissimilar metal additive manufacturing, characterized in that, It is applied to a dual-gun electric arc additive manufacturing system, wherein the dual-gun electric arc additive manufacturing system adopts a solid-liquid interface additive manufacturing layout with steel in the front and aluminum in the back; The method includes the following steps: The current process parameters are obtained and additive manufacturing is performed based on the current process parameters. The process parameters include welding torch spacing, aluminum welding torch current and aluminum welding torch voltage. Interface temperature field data is obtained during the manufacturing process, and temperature characteristic parameters are determined based on the interface temperature field data. The temperature characteristic parameters are input into the prediction model to obtain the interface brittle phase characteristics; based on the interface brittle phase characteristics, the interface bonding strength is determined. In response to the interface bonding strength not meeting the preset threshold, the reverse adjustment model is invoked to determine the adjustment amount of the process parameters; Adjust the additive manufacturing process parameters based on the aforementioned process parameter adjustment amount.

2. The method for interface control in steel-aluminum dissimilar metal additive manufacturing according to claim 1, characterized in that, The temperature characteristic parameters include the average cooling rate and the high-temperature residence time.

3. The method for interface control in steel-aluminum dissimilar metal additive manufacturing according to claim 1 or 2, characterized in that, The prediction model includes a first neural network and a second neural network; The temperature characteristic parameters are input into the prediction model to obtain the brittle phase characteristics of the interface, including: The average cooling rate is input into the first neural network, which is used to predict the composition type of the brittle phase at the interface based on the average cooling rate. The high-temperature residence time is input into the second neural network, which is used to predict the layer thickness of the brittle phase at the interface based on the high-temperature residence time.

4. The method for interface control in steel-aluminum dissimilar metal additive manufacturing according to claim 3, characterized in that, The prediction model also includes a third neural network; Based on the aforementioned brittle phase characteristics of the interface, the interfacial bonding strength is determined, including: The composition type and layer thickness of the brittle phase at the interface are input into the third neural network, which is used to predict the interfacial bonding strength based on the composition type and layer thickness of the brittle phase at the interface.

5. The method for interface control in steel-aluminum dissimilar metal additive manufacturing according to claim 4, characterized in that, The inputs to the reverse adjustment model include the current process parameters, temperature characteristic parameters, and interface bonding strength, and the output is the adjustment amount of the process parameters.

6. The method for interface control in steel-aluminum dissimilar metal additive manufacturing according to claim 1, characterized in that, Based on the aforementioned process parameter adjustment amount, the additive manufacturing process parameters are adjusted, including: The optimized process parameters are calculated based on the adjustment amount of the process parameters. Boundary verification is performed on the optimized parameters to ensure that the parameters are within the preset feasible range. The parameter adjustment is then performed through the dual-gun electric arc additive manufacturing system.

7. The method for interface control in steel-aluminum dissimilar metal additive manufacturing according to claim 3, characterized in that, The first neural network is a 3-layer backpropagation neural network, specifically including a 16-dimensional first hidden layer, an 8-dimensional second hidden layer, and a 4-dimensional third hidden layer; The second neural network is a 3-layer backpropagation neural network, specifically consisting of a 16-dimensional fourth hidden layer, an 8-dimensional fifth hidden layer, and a 4-dimensional sixth hidden layer.

8. The method for interface control in steel-aluminum dissimilar metal additive manufacturing according to claim 4, characterized in that, The third neural network is a 3-layer backpropagation neural network, specifically consisting of a 16-dimensional seventh hidden layer, an 8-dimensional eighth hidden layer, and a 4-dimensional ninth hidden layer.

9. The method for interface control in steel-aluminum dissimilar metal additive manufacturing according to claim 5, characterized in that, The backpropagation model is a 3-layer backpropagation neural network, specifically consisting of a 16-dimensional 10th hidden layer, an 8-dimensional 11th hidden layer, and a 4-dimensional 12th hidden layer.

10. A steel-aluminum dissimilar metal additive manufacturing interface control device, used to implement the method described in any one of claims 1-9, characterized in that, include: The data acquisition unit is used to acquire the current process parameters and perform additive manufacturing based on the current process parameters. During the manufacturing process, it acquires interface temperature field data and determines temperature characteristic parameters based on the interface temperature field data. The feature parameter prediction unit is used to input the temperature feature parameters into the prediction model to obtain the interface brittle phase characteristics; Based on the brittle phase characteristics of the interface, the interfacial bonding strength is determined; The process parameter adjustment unit is used to respond to the interface bonding strength not meeting the preset threshold by calling the reverse adjustment model, determining the process parameter adjustment amount, and adjusting the additive manufacturing process parameters based on the process parameter adjustment amount.

Images