Automatic welding method for oxygen lance nozzle copper-steel dissimilar metal

Abstract

The application discloses an oxygen lance nozzle copper-steel dissimilar metal automatic welding method, and belongs to the field of advanced manufacturing technology of key components of metallurgical equipment. The method comprises the following steps: establishing and calibrating a welding process digital twin based on a physical model and historical data, extracting multi-source heterogeneous features, adopting a channel-time double attention multi-modal fusion network to perform real-time state diagnosis and defect early warning on the features, taking a strategy map pre-generated by the digital twin as a benchmark, using an online reinforcement learning agent to realize adaptive closed-loop regulation and control of welding parameters, and continuously updating a knowledge base through an incremental learning mechanism. The application adopts a gas tungsten arc welding process, realizes offline global optimization of the process through digital twinning and deep reinforcement learning, realizes accurate control of a dynamic process in combination with multi-modal perception and online learning, and improves the consistency of dissimilar metal welding joint quality, process adaptive capacity and system autonomous evolution level, so that the strict manufacturing requirements of key components of high-end metallurgical equipment are met.

Description

Technical Field

[0001] This invention relates to the field of advanced manufacturing technology for key components of metallurgical equipment, specifically to an automated welding method for dissimilar metals such as copper and steel in oxygen lance nozzles. Background Technology

[0002] The oxygen lance nozzle is a core component of converter steelmaking. Its internal pipes supply oxygen into the furnace, while the external pipes are water circulation pipes. It operates in an extremely harsh environment, enduring high temperatures, high pressures, rapid heating and cooling, and slag erosion. The head of the oxygen lance nozzle is made of pure copper casting, while the rear is welded from copper to 20mm steel pipe or austenitic stainless steel. Due to significant differences in physical and metallurgical properties between copper and steel—for example, pure copper has a thermal conductivity more than 7 times greater than steel at 20℃ and more than 11 times greater at 1000℃, their melting points differ by approximately 500℃, and their coefficients of linear expansion differ by about 30%—and their poor miscibility in the liquid state, the welded joints of the oxygen lance nozzle exhibit significant residual deformation and stress, making them highly susceptible to cracking and causing obvious macroscopic segregation, hindering the fusion of the base material and filler material.

[0003] The main challenges in welding dissimilar metals like copper and steel are as follows: First, hot cracking is prone to occur in the weld. Copper segregates at grain boundaries, forming a low-melting-point eutectic structure. Furthermore, the significant difference in linear expansion coefficients between copper and steel generates substantial residual stress, further inducing hot cracking. Second, significant macroscopic segregation exists. In the high-temperature liquid-phase miscible region, large unstable miscible zones exist on both the copper and steel sides. The significant difference in melting points and thermal conductivity between copper and steel leads to liquid phase separation under rapid cooling conditions. The supercooled iron-copper mixture separates into iron and copper droplets, with the copper-rich phase solidifying later, resulting in significant macroscopic segregation. Third, controlling the melt ratio is difficult. Copper alloys have a lower melting point and higher thermal conductivity, making it difficult for the melting and solidification processes of copper and steel to occur simultaneously during welding. Due to its lower melting point and higher thermal conductivity, the copper side often melts before the steel side, making welding difficult. Gas metal arc welding (GMAW) has become the preferred process for welding dissimilar metals such as copper and steel due to its concentrated heat input, good controllability of the molten pool, and strong adaptability. However, this process is extremely sensitive to parameter fluctuations, and key issues such as hot cracking, lack of fusion, and significant macroscopic segregation control still need to be addressed.

[0004] Traditional welding methods heavily rely on operator experience, involving manual observation and parameter adjustments, resulting in poor quality stability and low efficiency. While existing automated welding equipment can perform welding according to fixed programs, it essentially still involves repetitive operations based on preset, fixed parameter trajectories. This method lacks real-time perception and intelligent decision-making capabilities during the welding process, and cannot effectively address workpiece assembly errors, beveling deviations, fluctuations in base material composition, and dynamic disturbances during welding. When operating conditions change, defects such as incomplete penetration, undercut, porosity, or cracks are easily generated, leading to fluctuations in product yield and high rework rates. Furthermore, conventional radiographic testing has limitations in assessing the performance of welded joints, making it difficult to comprehensively evaluate welding quality.

[0005] Furthermore, existing technologies have long process optimization cycles, requiring repeated process experiments, and knowledge accumulation relies on personal experience, making it difficult to solidify and reuse knowledge. The welding process is a complex dynamic process involving multi-physics coupling and multi-parameter interaction, but existing methods lack in-depth mining and utilization of massive amounts of process data, failing to achieve a fundamental shift in welding technology from experience-driven to data- and model-driven approaches.

[0006] To address the aforementioned issues, there is an urgent need for an automated welding method for dissimilar metals such as copper and steel using an oxygen lance nozzle, which can solve the problems associated with traditional methods. Summary of the Invention

[0007] The purpose of this invention is to provide an automated welding method for dissimilar metals, copper and steel, using oxygen lance nozzles. This method achieves offline global optimization of the process through digital twins and deep reinforcement learning, and combines multimodal perception and online learning to achieve precise dynamic process control. This improves the consistency of dissimilar metal welded joint quality, process adaptability, and system autonomous evolution level, meeting the stringent manufacturing requirements of key components for high-end metallurgical equipment.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows: An automated welding method for dissimilar metals, copper and steel, using an oxygen lance nozzle includes: Step 1: Establish a digital twin model of the welding process, build and train a welding strategy evaluation network based on the digital twin model, and use deep reinforcement learning algorithms to optimize the strategy, generating a welding strategy map and an initial strategy evaluation network. Step 2: Simultaneously acquire visual, acoustic emission, spectral, and temperature signals during the welding process using a multi-sensor system integrated into the welding torch, and extract real-time features from the signals; Step 3: Input the extracted feature vector into the preset multimodal fusion deep learning network and output the welding state diagnosis vector in real time. The welding state diagnosis vector includes the state classification probability distribution and the defect probability. Step 4: Based on the welding strategy map and starting from the initial strategy evaluation network, construct an online reinforcement learning agent; the agent receives the welding state diagnosis vector and outputs welding parameter adjustment instructions to achieve closed-loop control of the welding process; Step 5: Collect data from the entire welding process and incrementally learn and update the digital twin model, the initial strategy evaluation network, and the multimodal fusion deep learning network.

[0009] Further, in step 1, a digital twin model of the welding process is established. Based on the digital twin model, a welding strategy evaluation network is established and trained. A deep reinforcement learning algorithm is then used to optimize the strategy, generating a welding strategy map and an initial strategy evaluation network. Specifically: Step 101: Establish a finite element digital twin model that couples heat conduction, elastoplastic stress and strain, and microstructure evolution. The digital twin model includes heat conduction control equations, welding heat source model, stress and strain analysis module, and microstructure prediction module. Step 102: Calibrate and verify the digital twin model using historical welding data; Step 103: Conduct virtual welding experiments based on the calibrated model to generate training data and train the welding strategy evaluation network. The input of the welding strategy evaluation network is the welding parameter sequence, and the output is the predicted quality index. Step 104: Using the policy evaluation network as the environment model, run the near-end policy optimization algorithm for deep reinforcement learning to obtain the welding policy map.

[0010] Further, in step 103, the welding strategy evaluation network includes an input layer, a temporal feature encoder, and a quality index regression layer. The temporal feature encoder consists of three fully connected layers, and the quality index regression layer is a fully connected layer.

[0011] Further, in step 2, the multi-sensor system includes: The structured light vision sensor has its laser projector and high-speed camera mounted at an angle to the side and rear of the welding torch to acquire the three-dimensional morphology of the molten pool and weld surface. An acoustic emission sensor is installed on the steel gun body at a predetermined distance from the center line of the weld seam using a high-temperature coupling agent; An arc spectrometer probe is mounted on the side of the welding torch, with the focus aligned with the arc region between the tip of the welding wire and the molten pool. An infrared temperature sensor is aimed at the steel-side base material area near the copper-steel interface. The synchronous acquisition unit is used to synchronously acquire data from the sensors.

[0012] Furthermore, in step 2, real-time feature extraction is performed on the signal, specifically as follows: Extract the maximum width of the molten pool, the length of the molten pool, the trailing angle, and the weld reinforcement from the structured light image; The root mean square energy, peak frequency, and specific frequency band energy ratio extracted from the acoustic emission signal; The intensity ratio of copper and iron spectral lines and the intensity of oxygen spectral lines were extracted from the electric arc spectrum. Temperature and instantaneous cooling rate extracted from infrared signals; Instantaneous current, voltage, and welding speed obtained from the welding controller.

[0013] Furthermore, in step 3, the multimodal fusion deep learning network is a channel-temporal dual attention network, the structure of which includes, in sequence: Multi-branch one-dimensional convolution feature extraction module: Groups the input feature vectors according to sensor type, inputs multiple independent one-dimensional convolution branches for feature extraction, and concatenates the outputs of each branch; Channel attention module: Performs global average pooling on the concatenated features and generates channel attention weights through a fully connected layer to reweight the features; Temporal Attention LSTM Module: Input the weighted features into the bidirectional LSTM layer, calculate the attention weights of the hidden state at the last time step and the hidden states at each historical time step, generate a context vector, and concatenate the context vector with the hidden state at the last time step. Multi-task output module: Based on the concatenated features, after sharing a fully connected layer, the state classification probability distribution and defect probability are output through the state classification head and defect probability prediction head, respectively.

[0014] Furthermore, each one-dimensional convolutional branch in the multi-branch one-dimensional convolutional feature extraction module includes two one-dimensional convolutional layers and two max pooling layers; the first convolutional layer has a kernel size of 5 and a filter count of 16; the second convolutional layer has a kernel size of 3 and a filter count of 32; and each pooling layer has a pooling size of 2.

[0015] Furthermore, in step 5, data from the entire welding process is collected, and incremental learning and updates are performed on the digital twin model, the initial strategy evaluation network, and the multimodal fusion deep learning network. Specifically: The elastic weight consolidation method is used to fine-tune the multimodal fusion deep learning network to prevent it from catastrophically forgetting knowledge of old tasks. For digital twin models, complete data from each welding operation is used as new calibration and verification data, stored in the database, and calibration and verification are performed periodically. For the initial policy evaluation network, complete data from each weld is added to the training dataset as sample pairs, and the initial policy evaluation network is fine-tuned periodically using the training set.

[0016] In summary, the present invention has at least one of the following beneficial technical effects: 1. Achieving comprehensive intelligentization of the gas metal arc welding (GMAW) process. Addressing the challenges of hot cracking sensitivity and brittle phase formation in copper-steel dissimilar metal welding, this invention deeply integrates digital twins, multi-source sensing, deep learning, and reinforcement learning algorithms to construct a complete intelligent decision-making chain, from offline global optimization and online real-time diagnosis to adaptive closed-loop control. The system can simulate, perceive, understand, and autonomously optimize the welding process, replacing the traditional trial-and-error model relying on human experience and elevating welding technology to a new data- and model-driven stage.

[0017] 2. Possesses strong adaptive and anti-interference capabilities for gas metal arc welding (GMAW). Through real-time fine-tuning of welding parameters by an online reinforcement learning agent, the system can automatically compensate for interference caused by assembly errors, changes in weld gap, and fluctuations in heat accumulation, always maintaining the welding process in an optimal or near-optimal state. This dynamic adjustment capability significantly improves process robustness and ensures welding quality stability under complex and variable working conditions.

[0018] 3. Significantly improves the consistency and reliability of gas metal arc welding (GMAW) quality. Real-time status diagnosis based on a multimodal fusion network can accurately identify the microscopic state of the molten pool and provide early warnings of defects. Combined with online closed-loop control, it can effectively prevent the occurrence of various welding defects. This method reduces reliance on operator skills, enabling different batches of products to achieve high-standard and consistent welding quality, reducing scrap rates and rework costs.

[0019] 4. An evolutionary process knowledge system has been formed. Through the incremental learning and model update mechanism of the welding knowledge base, new data generated from each welding practice can be used to optimize the digital twin model, strategy network, and diagnostic model, enabling the system performance to continuously evolve with the accumulation of production data. This achieves the digital accumulation, autonomous iteration, and permanent inheritance of the company's core process knowledge.

[0020] 5. Improved the overall efficiency of gas metal arc welding (GMAW) process development and production. The combination of offline digital twins and deep reinforcement learning enables rapid optimization of massive process parameters in a virtual space, significantly shortening the traditional process testing cycle. In actual production, intelligent adaptive control reduces downtime caused by quality fluctuations, thereby improving equipment utilization and production efficiency. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0023] like Figure 1 As shown, this invention provides an automated welding method for dissimilar metals, copper and steel, using an oxygen lance nozzle and employing a gas metal arc welding (GMAW) process, comprising: Step 1: Establish a digital twin model of the welding process, build and train a welding strategy evaluation network based on the digital twin model, and use deep reinforcement learning algorithms to optimize the strategy, generating a welding strategy map and an initial strategy evaluation network. Step 2: Simultaneously acquire visual, acoustic emission, spectral, and temperature signals during the welding process using a multi-sensor system integrated into the welding torch, and extract real-time features from the signals; Step 3: Input the extracted feature vector into the preset multimodal fusion deep learning network and output the welding state diagnosis vector in real time. The welding state diagnosis vector includes the state classification probability distribution and the defect probability. Step 4: Based on the welding strategy map and starting from the initial strategy evaluation network, construct an online reinforcement learning agent; the agent receives the welding state diagnosis vector and outputs welding parameter adjustment instructions to achieve closed-loop control of the welding process; Step 5: Collect data from the entire welding process and incrementally learn and update the digital twin model, the initial strategy evaluation network, and the multimodal fusion deep learning network.

[0024] This method targets the welding characteristics of dissimilar metals such as copper and steel, using gas metal arc welding (GMAW) as the process carrier to establish a mapping relationship between process parameters and welding quality.

[0025] The above steps will now be described in detail. In step 1, a digital twin model of the welding process is established. Based on the digital twin model, a welding strategy evaluation network is built and trained. Deep reinforcement learning algorithms are then used to optimize the strategy, generating a welding strategy map and an initial strategy evaluation network. Specifically: Step 101: For the characteristics of gas metal arc welding (GMAW), a finite element digital twin model is established that couples heat conduction, elasto-plastic stress-strain, and microstructure evolution. The digital twin model includes heat conduction control equations, a welding heat source model, a stress-strain analysis module, and a microstructure prediction module. The welding heat source model adopts a combination of a double ellipsoidal heat source and a Gaussian surface heat source to accurately simulate the arc heat flux density distribution in GMAW. Specifically: 1. Governing equations for heat conduction The welding temperature field is described by a transient nonlinear heat conduction equation. The evolution is as follows: ; In the formula, For material density, is the specific heat capacity, k is the thermal conductivity, both are functions of temperature T, and Q is the welding heat source input term; 2. Welding heat source model: To accurately simulate the arc heat flux density distribution, a combined model of a double ellipsoidal heat source and a Gaussian surface heat source is used, taking into account the deep and narrow bevel characteristics of the oxygen lance nozzle. For the front half of the double ellipsoidal heat source, the heat flux density... Represented as: ; in, For the effective welding heat source input, η is the arc thermal efficiency, and U and I are the welding voltage and current, respectively. a f b and c are the shape parameters of the heat source. The energy distribution coefficient for the front half of the ellipsoid. x , y , z The coordinates are .

[0026] 3. Stress-strain analysis: A thermo-elastoplastic constitutive model is used to calculate the transient thermal stress σ and strain caused by the non-uniform temperature field. : ; Where D is the temperature-dependent elasticity matrix. The thermal strain increment (α is the coefficient of linear expansion). The plastic strain increment is obtained by using the Mises yield criterion and related flow rules.

[0027] 4. Microstructure prediction: Based on continuous cooling transformation kinetics, the microstructure proportions of the heat-affected zone are calculated. The martensite transformation volume fraction is calculated using the superposition rule. : ; Where κ(T) is a temperature-dependent phase transition kinetic function.

[0028] Step 102: Calibrate and verify the digital twin model using historical welding data, specifically as follows: 1. Input calibration data: Select an array containing the actual welding process parameters with post-weld inspection results, the corresponding temperature cycle curves recorded by the infrared thermal imager, and metallographic photographs of the weld cross-section.

[0029] 2. Parameter Inversion: Key uncertainty parameters in the digital twin model are adjusted through optimization algorithms, mainly the parameters of the welding heat source model. a f b, c The parameters η and the boundary heat transfer coefficient are used to minimize the error between the simulated temperature cycle curve and the molten pool shape and the measured data. An error function E is defined as follows: ; in, and The simulation and experimental temperatures are respectively. These are the weighting coefficients.

[0030] 3. Model Validation: Using another set of uncalibrated measured data, compare the consistency between the simulation-predicted weld macromorphology and heat-affected zone width and the measured results. When the errors of key indicators are all less than the preset threshold (e.g., weld width error <5%), the model is considered calibrated and can be used for subsequent virtual experiments.

[0031] Step 103: Conduct virtual welding experiments based on the calibrated model to generate training data and train the welding strategy evaluation network. The input of the welding strategy evaluation network is the welding parameter sequence, and the output is the predicted quality index, specifically: 1. Virtual Experiment Data Generation: On the calibrated digital twin model, large-scale virtual welding simulations are conducted within a broad welding parameter space according to the experimental design methodology. Each simulation records the complete parameter sequence and corresponding output results.

[0032] 2. Dataset Construction: The input feature vector for each data sample is a normalized sequence of welding parameters. The output target vector contains multiple quality indicators obtained from simulation calculations.

[0033] 3. A detailed explanation of the structure and training of the welding strategy evaluation network: The welding strategy evaluation network employs a deep fully connected neural network based on an encoder structure to capture the temporal dependencies and high-order feature interactions in the welding parameter sequence. Its network structure, from top to bottom, is as follows: (1) Input layer: Input data: A complete sequence of welding parameters, describing the changes of all key parameters over time in a virtual welding task. This sequence is represented as a matrix. In the formula, This represents the total number of time steps for discretizing the welding process. The welding parameters considered include: peak current, base current, pulse frequency, welding speed, wire feed speed, oscillation amplitude, oscillation frequency, arc voltage, and shielding gas flow rate for gas metal arc welding (GMAW). All parameters are normalized before input.

[0034] (2) Temporal feature encoder (consisting of three fully connected layers): First coding layer: Receives the vector (dimension) after flattening the input matrix P. This layer contains 256 neurons and uses the ReLU activation function. Its function is to perform high-order feature extraction and dimensionality compression.

[0035] The second coding layer contains 128 neurons and uses the ReLU activation function. It further refines abstract features strongly correlated with welding quality.

[0036] The third coding layer contains 64 neurons and uses the ReLU activation function. This layer outputs a highly condensed, fixed-length feature vector. This vector encodes global information for the entire sequence of welding parameters.

[0037] (3) Quality index regression layer (output layer): This layer is a fully connected layer, with the input being the encoded feature vector z, and the output dimension being... That is, the number of quality indicators to be predicted.

[0038] This layer does not use an activation function; that is, it uses linear output to suit the regression task.

[0039] The output vector is Each of them This corresponds to a normalized predicted value for a quality indicator.

[0040] In addition, this invention provides typical quality indicators. include: Maximum deformation :: The maximum normal displacement of the workpiece after cooling; Interfacial residual stress Peak residual stress perpendicular to the weld direction at the copper-steel interface; Crack sensitivity index A comprehensive index calculated based on the stress field, brittle phase fraction, and heat-affected zone width obtained from simulation is used to assess the risk of hot and cold cracking. Virtual penetration depth H: The penetration depth at the root of the weld.

[0041] Next, the training process and working principle will be explained in detail: (1) Training process Loss function: The mean squared error loss function is used to measure the difference between the network's predicted values ​​and the true values ​​of the digital twin simulation, which is: ; In the formula, N is the training batch size. This is the network prediction value. For digital twin simulation of truth value; Optimization: Minimize using the Adam optimizer Early stopping is used during training; training is terminated when the loss on the validation set no longer decreases, in order to obtain a model with the best generalization ability.

[0042] (2) Working principle After training, the weights within the network determine the mapping function from the multidimensional parameter sequence space to the quality metric space. In deep reinforcement learning training, whenever the agent tries a sequence of parameters (actions)... The strategy evaluation network is invoked to calculate the prediction quality. The reward function R in reinforcement learning is based on... Computation. The efficient forward propagation of the network enables the agent to perform millions of policy trials and evaluations in a short period of time, thereby achieving global optimization.

[0043] Step 104: Using the policy evaluation network as the environment model, run the near-end policy optimization algorithm for deep reinforcement learning to obtain the welding policy map, specifically: Using the trained policy evaluation network as a fast environment model, a deep reinforcement learning algorithm is run to find the optimal welding strategy.

[0044] 1. Definition of reinforcement learning problem: state : Normalized coding of the current welding process, including the proportion of completed welds and the baseline of the main parameters for the current stage.

[0045] action Adjustments to key welding parameters, such as , Its scope is set based on the feasibility of the process.

[0046] award The reward function R is calculated by the policy evaluation network based on the predicted quality index after welding is completed. ; In the formula, These are the penalty weighting coefficients for each indicator, all of which are positive numbers. The higher the reward value, the better the overall quality.

[0047] Strategy : A neural network defined by parameter θ, with input state Output action The probability distribution.

[0048] 2. Algorithm Implementation: The near-end policy optimization algorithm is adopted. A policy network is maintained. and a value network In a virtual environment, the agent interacts with the environment according to the current policy, generating trajectory data. Using this data, the policy network parameters θ are updated by maximizing an objective function with pruning terms. ; in, This represents the probability ratio between the old and new strategies. The advantage function estimate is obtained through the value network. And obtained from the generalized dominance estimation algorithm, To prune hyperparameters, the value network updates by minimizing the mean squared error between its predicted values ​​and actual returns.

[0049] 3. Strategy Graph Generation: After sufficient training, the agent learns an optimal parameter adjustment strategy from any given initial state to the end of welding. This strategy is discretized and sampled in the state space to form a welding strategy graph Π0 covering typical working conditions. This graph is essentially a lookup table or function mapping, which can provide recommended baseline parameters and allowable adjustment ranges for a given welding stage and conditions.

[0050] In step 2, visual, acoustic emission, spectral, and temperature signals from the welding process are simultaneously acquired using a multi-sensor system integrated into the welding torch, and real-time feature extraction is performed on the signals. Specifically: 1. First, the integration and deployment of multi-sensor systems will be introduced: The following sensors are mechanically rigidly connected to the gas metal arc welding torch holder at the end of the welding robot, ensuring a constant relative position between them and the welding torch. Heat insulation, splash prevention, and electromagnetic shielding measures are implemented, specifically: (1) Structured light vision sensor: Components: A semiconductor laser line projector with a wavelength of 850 nanometers and a high-speed CMOS industrial camera equipped with a corresponding bandpass filter.

[0051] Installation location: The laser projector is positioned at a pre-calibrated fixed angle to the camera, behind the welding torch, approximately 30° above and to the side of the welding torch when viewed along the welding direction. The laser plane is perpendicular to the welding direction and projected onto the workpiece surface approximately 5 mm in front of the welding torch.

[0052] Function: The laser stripes deform the surface of the molten pool and the nearby workpiece. The camera captures the image of these stripes and uses it to reconstruct the three-dimensional morphology of the molten pool and the solidified weld behind it.

[0053] (2) Acoustic emission sensor: Components: A wideband piezoelectric acoustic emission sensor with a resonant frequency of 150kHz and a preamplifier.

[0054] Installation location: The two acoustic emission sensors are rigidly coupled via high-temperature couplant to the upstream and downstream positions of the steel lance body section of the oxygen lance nozzle, approximately 50 mm from the weld centerline. The sensor axes are perpendicular to the workpiece surface.

[0055] Function: To collect transient elastic stress wave signals generated during the welding process due to plastic deformation, crack initiation and propagation, and molten pool fluctuations.

[0056] (3) Arc spectrometer probe: Components: Fiber-optic guided miniature spectrometer probe, covering a spectral range of 350 nm to 850 nm.

[0057] Installation location: The probe is installed on the side of the welding torch, with its receiving lens axis at an angle of approximately 45° to the welding wire axis, and the focal point aligned with the arc area between the welding wire tip and the molten pool.

[0058] Function: To collect continuous spectra and characteristic atomic / ion emission lines of electric arc plasma.

[0059] (4) Infrared temperature measurement point sensor: Component: High-speed infrared thermometer with a response wavelength in the mid-infrared band.

[0060] Installation location: The sensor is positioned near the copper-steel interface, approximately 3 mm from the center line of the weld, on the steel side of the base material, and the measurement point is fixed.

[0061] Function: To monitor the thermal cycle curve of this critical location in real time, especially the peak temperature and cooling rate.

[0062] (5) Multi-channel synchronous data acquisition system All sensor signals are connected to a central synchronous acquisition unit. This unit uses a high-precision PXle timing module to uniformly generate sampling clocks and trigger signals for all channels, ensuring that the visual, acoustic emission, spectral, temperature, and process parameter data streams have a unified time base, with a time synchronization accuracy better than 1 microsecond. All raw data streams are timestamped uniformly. 2. Next, we will provide a detailed introduction to the specific feature extraction of sensor data: (1) Extracting morphological features of the molten pool and weld from structured light images: For each frame of laser stripe image, threshold segmentation and skeleton extraction are performed to calculate the key geometric parameters of the molten pool region: Maximum width of molten pool The maximum lateral pixel distance of the stripe deformation portion in the molten pool area is converted to physical size (millimeters) after calibration. molten pool length : Maximum pixel distance along the welding direction, converted to physical size (millimeters); Trailing corner : The angle (in degrees) between the tangent at the tail boundary of the molten pool and the normal to the workpiece surface; weld reinforcement : The height (in millimeters) of the stripe at the center line of the solidified weld area relative to the reference plane; (2) Extracting time-frequency domain features from acoustic emission signals: Perform a Fast Fourier Transform on the raw acoustic emission waveform for each time window (e.g., 10 milliseconds) and calculate: Root mean square energy The effective energy value of a signal within a window; Peak frequency : The frequency point (kHz) with the highest power spectral density; Energy ratio of a specific frequency band The ratio of energy to total energy within the 100kHz to 200kHz frequency band. This frequency band is highly correlated with microcrack formation. (3) Extracting metallurgical reaction characteristics from electric arc spectroscopy: For the spectral curves of each acquisition cycle, identify the intensity of characteristic spectral lines of copper atoms (e.g., 510.5 nm) and characteristic spectral lines of iron atoms (e.g., 404.5 nm). and ; Calculate the intensity ratio of copper-iron spectral lines This ratio is related to the degree of copper-steel mixing in the molten pool; Monitoring the intensity of oxygen atom spectral lines (e.g., 777 nm) As an indicator of atmospheric disturbance, this indicator is directly related to the purity and flow stability of the protective gas and is used to judge the protection effect in real time. (4) Extracting thermal cycling characteristics from infrared signals: Calculate the current temperature T(t) in real time; Calculate the instantaneous cooling rate Calculate the average cooling rate within a specific temperature range (e.g., 800℃ to 500℃). .

[0063] Finally, the feature vectors are encapsulated and output. At the end of each acquisition cycle, all the extracted feature values ​​are combined with the current process parameter values ​​from the controller. The combination of these features forms a real-time feature vector, which is: ; This feature vector fully reflects the metallurgical reaction at the copper-steel dissimilar metal interface and the dynamic behavior of the molten pool during gas metal arc welding (GMAW). This vector is encapsulated into a data packet and transmitted in real-time via high-speed industrial Ethernet to the real-time computing platform used in step 3. Simultaneously, all timestamped raw data and feature vectors are stored in a historical database, which will be used for the calibration and verification of the digital twin model in step 1 and for updating the knowledge base in step 5.

[0064] In step 3, the extracted feature vector is input into a preset multimodal fusion deep learning network, which outputs a welding state diagnosis vector in real time. This welding state diagnosis vector includes a state classification probability distribution and a defect probability, specifically: First, the input and output of the multimodal fusion deep learning network are determined, with the input being the feature vector. And a corresponding historical database, containing a large number of feature sequences and their labeled welding status tags and final defect records; the output is a real-time welding status diagnostic vector. Where C is a probability distribution vector representing the state category. It is a scalar that represents the overall probability of a major welding defect occurring within a specific future time window.

[0065] Secondly, the model structure is described in detail. The model consists of four core modules: a multi-branch one-dimensional convolutional feature extraction module, a channel attention module, a temporal attention LSTM module, and a multi-task output module. Each module is described in detail below: 1. Multi-branch one-dimensional convolution feature extraction module The multi-branch one-dimensional convolutional feature extraction module groups the input features according to their physical meaning, extracts features through multiple independent convolutional branches, and then concatenates them. Specifically: (1) Branching: Divide the F-dimensional input features into K groups according to sensor type (K=4 in the example): Branch 1: Morphological features (such as weld pool width, length, trailing angle, weld reinforcement height), totaling 4 dimensions.

[0066] Branch 2: Acoustic emission characteristics (such as root mean square energy, peak frequency, and energy ratio in a specific frequency band), totaling 3 dimensions.

[0067] Branch 3: Spectral characteristics (such as the ratio of copper to iron spectral line intensity, and the intensity of oxygen spectral line), totaling 2 dimensions.

[0068] Branch 4: Thermal cycle and process parameter characteristics (such as temperature, cooling rate, current, voltage, welding speed), a total of 5 dimensions.

[0069] The structure of each branch (each branch is independent and has the same structure): First convolutional layer: one-dimensional convolution, kernel size = 5, number of filters = 16, stride = 1, using ReLU activation function.

[0070] First pooling layer: one-dimensional max pooling, pooling size = 2, step size = 2.

[0071] Second convolutional layer: one-dimensional convolution, kernel size = 3, number of filters = 32, stride = 1, using ReLU activation function.

[0072] Second pooling layer: one-dimensional max pooling, pooling size = 2, step size = 2.

[0073] For the k-th branch, the output feature map of the l-th layer Calculation as follows ; Where * denotes a one-dimensional convolution operation, and For trainable weights and biases; Each branch outputs an abstracted feature sequence. The output features of all branches are concatenated along their feature dimensions to form a fused multidimensional feature sequence. ,in is the number of time steps after pooling, and D is the total feature dimension.

[0074] 2. Channel Attention Module (1) Design Principles In fusion features The network introduces a channel attention mechanism, which allows the network to adaptively learn and strengthen the weights of feature channels that are crucial to the current welding status judgment, while suppressing the influence of irrelevant or noisy channels.

[0075] (2) Structural details: Its structure includes a global average pooling layer and two fully connected layers, and their specific operations are described below: right In the time dimension Perform global average pooling to generate a channel description vector. .

[0076] The channel attention weight vector is generated by passing u through a fully connected network with a single hidden layer. .

[0077] ; in, δ is the trainable parameter, σ is the ReLU activation function, and σ is the Sigmoid activation function, which restricts the weights to between 0 and 1.

[0078] Original fusion features With attention weight Multiplying by channel one by one yields the weighted feature sequence. .

[0079] 3. Temporal Attention LSTM Module (1) Design principle: Changes in welding state are often closely related to key events (such as arc initiation, arc termination, and oscillation inflection point). Standard LSTM treats all time step information equally. This module introduces a temporal attention mechanism on the basis of LSTM, enabling it to focus on the time step that has the most historical influence on diagnosing the current state.

[0080] (2) Structural details: Weighted feature sequence Input a bidirectional LSTM layer. This layer contains two LSTMs, one for forward and one for backward, each with 64 hidden units. They process the sequence from two directions respectively, and finally concatenate the forward and backward hidden states at each time step to obtain a hidden state sequence containing bidirectional context information. .

[0081] Calculate temporal attention on the LSTM output. First, calculate the current last time step. Hidden state Hidden states at all time steps in the sequence correlation score : ; in, These are trainable parameters. The scores are then normalized to attention weights using the Softmax function. : ; Calculate the context vector As a weighted summary of the entire input sequence: ; context vector Hidden state at the last time step The data are then concatenated to form the final temporal feature representation. .

[0082] 4. Multi-task output module (1) Design principle: Welding condition diagnosis includes two closely related tasks: classifying the current condition and regressing and predicting future risks. Sharing underlying features and processing them through independent output heads can improve the model's generalization ability and efficiency.

[0083] (2) Structural details: Will Input a shared fully connected layer (64 neurons, ReLU activated).

[0084] Then it splits into two independent output heads: ① State Classification Head: A fully connected layer with a Softmax activation function, the number of output neurons equals the number of state categories. (For example, 4 categories: excellent, overheated, poorly fused, poorly formed). Its output C represents the probability distribution of belonging to each category.

[0085] ② Defect probability prediction head: A fully connected layer with a Sigmoid activation function, outputting a single scalar P d ∈[0,1] represents the probability of defects such as pores or microcracks appearing within the next 2 seconds.

[0086] Next, the training process of the training model will be explained in detail: 1. Construct the training dataset (1) Data source: The entire dataset comes from the historical welding process database of gas metal arc welding accumulated in step 2, which includes welding process data under different combinations of process parameters and the corresponding post-weld quality inspection results. Each sample consists of a continuous real-time feature vector sequence with a time length of T seconds. constitute.

[0087] (2) Sample labeling: ① Status Labeling: Senior welding process experts were hired to replay welding videos (from high-speed cameras) and all sensor data for each sample within the corresponding time period. The welding status at the sample's center time t was manually labeled and divided into... One of the pre-ordered categories.

[0088] ②Defect Labeling: Based on the non-destructive testing report (such as X-ray inspection) after the welding is completed, if a defect occurs within 2 seconds after the end of the corresponding time period for the sample, then the sample will be labeled with a defect label. Set it to 1, otherwise set it to 0.

[0089] (3) Data set partitioning: The total dataset is randomly divided into training set, validation set and test set in a ratio of 7:2:1.

[0090] 2. Model training process: (1) Loss function design: The multi-task loss function is adopted, which is a weighted sum of classification loss and regression loss.

[0091] State classification tasks use cross-entropy loss: ; The defect prediction task uses binary cross-entropy loss: ; Total loss: ,in and This is a hyperparameter used to balance the importance of the two tasks.

[0092] (2) Training process: Using the Adam optimizer, the initial learning rate was set to 0.001; Set the batch size to 64; Perform forward and backward propagation on the training set to update the network weights; After each training cycle, the loss is calculated on the validation set. If the validation loss does not decrease for several consecutive cycles, the learning rate is reduced or training is stopped early to prevent overfitting. Finally, the model parameters that showed the best overall performance on the validation set were selected as the trained diagnostic model.

[0093] In step 4, based on the welding strategy map and starting from the initial strategy evaluation network, an online reinforcement learning agent is constructed. The agent receives welding state diagnostic vectors and outputs welding parameter adjustment instructions to achieve closed-loop control of the welding process, specifically: First, the input and output of step 4 are determined. The input is: The welding strategy map Π0 and the initial parameters of the strategy evaluation network from step 1; Real-time welding status diagnostic vector from step 3 ; Current welding stage information (such as the percentage of welds completed, current layer number) and preset welding task objectives (such as highest quality, highest efficiency, etc.). Its output is: parameter adjustment instructions sent to the welding robot control system and welding power source, which are given in the form of fine-tuning increments, for example... and wait.

[0094] The specific operating procedures will be explained in detail below: 1. Agent initialization: Create an online reinforcement learning agent with a core policy network. and value network The initial weights are inherited from the final policy network obtained through deep reinforcement learning training in step 1. and value network The weighting of these factors provides a high-performance starting point for online learning, avoiding the initial high-risk period associated with learning from random strategies.

[0095] The welding strategy map Π0 is loaded into the agent's decision cache as a benchmark reference strategy for parameter adjustment.

[0096] 2. Real-time state construction: In each control cycle t, the agent receives a diagnostic vector from step 3. Coding with the current welding stage (A vector containing information such as the percentage of completed welds and the current layer number) are concatenated to form the complete state observation vector of the reinforcement learning agent. : ; C provides the classification judgment for the current microstate. The near-term risks have been quantified. It provides information on the macro-level process.

[0097] The following section explains online decision-making and learning based on the proximal policy optimization algorithm: 1. Definition of Action Space: action Defined as a normalized adjustment to key welding parameters. For example: ; in, and These are the physical adjustments for peak current and welding speed, respectively. and The preset adjustment standard deviation is used for normalization. Action values ​​are limited to... Within the specified range, ensure that the adjustment range is within the process safety zone.

[0098] 2. Reward function design: reward function The design aims to guide the agent to maintain excellent welding conditions and avoid defects. Its computation relies on state diagnostic vectors. : ; First item Encourage agents to move their states closer to the "good" category. It is the probability value of the "excellent" category in the diagnostic vector; Second item Penalize high defect probability; Third item Encourage gradual parameter adjustments and avoid drastic oscillations; Fourth item This refers to stage-specific rewards, such as additional goal-oriented rewards or penalties at specific points (e.g., arc initiation, arc termination, copper-steel interface). Weighting coefficient. It is a positive number.

[0099] 3. Online strategy optimization: An online learning algorithm, employing a proximal policy optimization method, is used. During the welding process, the agent learns according to the current online policy. Select Action And execute.

[0100] After the action is performed, the welding system enters a new state. and receive a reward This empirical tuple It is stored in a fixed-size online experience replay buffer.

[0101] Periodically, a small batch of empirical data is sampled from the buffer to update the policy network and value network. PPO updates the policy network parameters θ by optimizing an alternative objective function with clipping terms. ; in, For probability ratio The advantage function is calculated through generalized advantage estimation. These are hyperparameters. The value network updates by minimizing the mean squared error between its predicted values ​​and the actual returns.

[0102] Ultimately, decision-making, implementation, and safety oversight are carried out, specifically as follows: 1. Decision fusion: The actions output by the agent. After inverse normalization, the physical adjustment amount is obtained. , These adjustments will be added to the current stage baseline parameters obtained from the welding strategy map Π0 to form the final parameter settings sent to the actuator.

[0103] 2. Safety Monitor: A separate safety monitoring rules module is set up. This module monitors in real time. Defect probability in And original sensing features (such as melt pool width). If Exceeding the safety threshold If a critical feature exceeds physical limits, the monitor will immediately trigger safety protocols, such as pausing welding or forcibly reverting parameters to the baseline value in Π0, to ensure process safety.

[0104] In step 5, data from the entire welding process is collected, and incremental learning and updates are performed on the digital twin model, the initial strategy evaluation network, and the multimodal fusion deep learning network. Specifically: The elastic weight consolidation method is used to fine-tune the multimodal fusion deep learning network to prevent it from catastrophically forgetting knowledge of old tasks. For digital twin models, complete data from each welding operation is used as new calibration and verification data, stored in the database, and calibration and verification are performed periodically. For the initial strategy evaluation network, complete data from each welding is added to the training dataset as sample pairs. The initial strategy evaluation network is fine-tuned periodically using the training set, so that the process knowledge base can continuously accumulate optimal process experience of gas metal arc welding in copper-steel dissimilar metal welding.

[0105] Embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0106] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0107] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0108] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0109] Contents not described in detail in this specification are prior art known to those skilled in the art. It is hereby indicated that the above description is intended to help those skilled in the art understand this invention, but does not limit the scope of protection of this invention. Any equivalent substitutions, modifications, improvements, or simplifications of the above descriptions that do not depart from the essential content of this invention fall within the scope of protection of this invention.

Claims

1. An automated welding method for dissimilar metals, copper and steel, using an oxygen lance nozzle, characterized in that: include: Step 1: Establish a digital twin model of the welding process, build and train a welding strategy evaluation network based on the digital twin model, and use deep reinforcement learning algorithms to optimize the strategy, generating a welding strategy map and an initial strategy evaluation network. Step 2: Simultaneously acquire visual, acoustic emission, spectral, and temperature signals during the welding process using a multi-sensor system integrated into the welding torch, and extract real-time features from the signals; Step 3: Input the extracted feature vector into the preset multimodal fusion deep learning network and output the welding state diagnosis vector in real time. The welding state diagnosis vector includes the state classification probability distribution and the defect probability. Step 4: Based on the welding strategy map and starting from the initial strategy evaluation network, construct an online reinforcement learning agent; The intelligent agent receives welding status diagnostic vectors and outputs welding parameter adjustment instructions to achieve closed-loop control of the welding process; Step 5: Collect data from the entire welding process and incrementally learn and update the digital twin model, the initial strategy evaluation network, and the multimodal fusion deep learning network.

2. The automated welding method for dissimilar metals (copper and steel) using an oxygen lance nozzle according to claim 1, characterized in that, In step 1, a digital twin model of the welding process is established. Based on the digital twin model, a welding strategy evaluation network is built and trained. Deep reinforcement learning algorithms are then used to optimize the strategy, generating a welding strategy map and an initial strategy evaluation network. Specifically: Step 101: Based on the characteristics of gas metal arc welding (GMAW) process, establish a finite element digital twin model that couples heat conduction, elastic-plastic stress-strain, and microstructure evolution. The digital twin model includes heat conduction control equations, welding heat source model, stress-strain analysis module, and microstructure prediction module. Step 102: Calibrate and verify the digital twin model using historical welding data; Step 103: Conduct virtual welding experiments based on the calibrated model to generate training data and train the welding strategy evaluation network. The input of the welding strategy evaluation network is the welding parameter sequence, and the output is the predicted quality index. Step 104: Using the policy evaluation network as the environment model, run the near-end policy optimization algorithm for deep reinforcement learning to obtain the welding policy map.

3. The automated welding method for dissimilar metals (copper and steel) using an oxygen lance nozzle according to claim 2, characterized in that, In step 103, the welding strategy evaluation network includes an input layer, a temporal feature encoder, and a quality index regression layer. The temporal feature encoder consists of three fully connected layers, and the quality index regression layer is a fully connected layer.

4. The automated welding method for dissimilar metals (copper and steel) using an oxygen lance nozzle according to claim 3, characterized in that, In step 2, the multi-sensor system includes: The structured light vision sensor, with its laser projector and high-speed camera mounted at an angle to the side and rear of the gas metal arc welding torch, is used to acquire the three-dimensional morphology of the molten pool and weld surface. An acoustic emission sensor is installed on the steel gun body at a predetermined distance from the center line of the weld seam using a high-temperature coupling agent; An arc spectrometer probe is mounted on the side of the welding torch, with the focus aligned with the arc region between the tip of the welding wire and the molten pool. An infrared temperature sensor is aimed at the steel-side base material area near the copper-steel interface. The synchronous acquisition unit is used to synchronously acquire data from the sensors.

5. The automated welding method for dissimilar metals (copper and steel) using an oxygen lance nozzle according to claim 4, characterized in that, In step 2, real-time feature extraction is performed on the signal, specifically as follows: Extract the maximum width of the molten pool, the length of the molten pool, the trailing angle, and the weld reinforcement from the structured light image; The root mean square energy, peak frequency, and specific frequency band energy ratio extracted from the acoustic emission signal; The intensity ratio of copper and iron spectral lines and the intensity of oxygen spectral lines were extracted from the arc spectrum. The intensity of the oxygen spectral line was used to monitor the purity and flow stability of the protective gas. Temperature and instantaneous cooling rate extracted from infrared signals; Instantaneous current, voltage, and welding speed obtained from the welding controller.

6. The automated welding method for dissimilar metals (copper and steel) using an oxygen lance nozzle according to claim 5, characterized in that, In step 3, the multimodal fusion deep learning network is a channel-temporal dual attention network, the structure of which includes: Multi-branch one-dimensional convolution feature extraction module: Groups the input feature vectors according to sensor type, inputs multiple independent one-dimensional convolution branches for feature extraction, and concatenates the outputs of each branch; Channel attention module: Performs global average pooling on the concatenated features and generates channel attention weights through a fully connected layer to reweight the features; Temporal Attention LSTM Module: Input the weighted features into the bidirectional LSTM layer, calculate the attention weights of the hidden state at the last time step and the hidden states at each historical time step, generate a context vector, and concatenate the context vector with the hidden state at the last time step. Multi-task output module: Based on the concatenated features, after sharing a fully connected layer, the state classification probability distribution and defect probability are output through the state classification head and defect probability prediction head, respectively.

7. The automated welding method for dissimilar metals (copper and steel) using an oxygen lance nozzle according to claim 6, characterized in that, Each one-dimensional convolutional branch in the multi-branch one-dimensional convolutional feature extraction module includes two one-dimensional convolutional layers and two max pooling layers; the first convolutional layer has a kernel size of 5 and a filter count of 16; the second convolutional layer has a kernel size of 3 and a filter count of 32; and each pooling layer has a pooling size of 2.

8. The automated welding method for dissimilar metals (copper and steel) using an oxygen lance nozzle according to claim 7, characterized in that, In step 5, data from the entire welding process is collected, and incremental learning and updates are performed on the digital twin model, the initial strategy evaluation network, and the multimodal fusion deep learning network. Specifically: The elastic weight consolidation method is used to fine-tune the multimodal fusion deep learning network to prevent it from catastrophically forgetting knowledge of old tasks. For digital twin models, complete data from each welding operation is used as new calibration and verification data, stored in the database, and calibration and verification are performed periodically. For the initial policy evaluation network, complete data from each weld is added to the training dataset as sample pairs, and the initial policy evaluation network is fine-tuned periodically using the training set.

Images