Copper-aluminum heterogeneous material laser welding mechanical property prediction model

Abstract

The invention relates to the field of laser welding quality monitoring, in particular to a copper-aluminum heterogeneous material laser welding mechanical property prediction model. Comprising the steps that a multi-source data acquisition module acquires laser power, scanning speed and the like in real time through a sensor array; the physical mechanism modeling module is based on a multi-physics field coupling finite element algorithm, utilizes a self-adaptive heat source unit to dynamically compensate copper-aluminum heat conductivity difference, and simulates temperature field and residual stress evolution; the data driving prediction module adopts a dual-channel deep neural network, a first channel inputs a process parameter vector, a second channel inputs a molten pool geometric feature and a thermal cycle vector, and prediction values of tensile strength, elongation and hardness are jointly output through a feature fusion layer; and the online decision feedback module compares the predicted strength with a safety threshold value, generates a process parameter correction instruction through Bayesian optimization, and drives an execution mechanism to perform closed-loop control. A physical mechanism and dynamic feature capture are fused, the problem of high cost of a traditional trial and error method is solved, and prediction precision and process stability are synchronously improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of laser welding quality monitoring, more particularly, the present application relates to a copper-aluminum heterogeneous material laser welding mechanical property prediction model. BACKGROUND

[0002] Copper-aluminum heterogeneous material welding structure is widely used in new energy battery module, power electronic heat sink and aerospace lightweight components, which has the advantages of high conductivity of copper and lightweight of aluminum. However, the physical properties of copper and aluminum are significantly different - the thermal conductivity of copper (398 W / m·K) is about 1.7 times that of aluminum (237 W / m·K), and the linear expansion coefficient difference is up to 40%, resulting in two major problems in the laser welding process: first, the energy absorption rate of high reflectivity copper material fluctuates (15%-30% deviation), which causes poor fusion; second, the intermetallic compound (IMC) grows excessively (thickness > 5 μm, joint ductility drops sharply), and non-uniform thermal expansion causes residual stress concentration. Traditional process development relies on trial and error experiments, with a single working condition cost of more than ten thousand yuan, and cannot establish a quantitative mapping relationship between process parameters and mechanical properties, which seriously restricts the industrial application of high-reliability joints;

[0003] The prior art has three bottlenecks:

[0004] Empirical trial and error method is inefficient: the response surface method based on orthogonal experiment needs at least 25 repeated experiments (more than 2 weeks), and cannot capture the correlation between molten pool dynamic behavior and IMC growth, and the optimization result is only suitable for a specific device working condition;

[0005] Pure physical simulation has insufficient accuracy: the finite element model (such as ANSYS) is difficult to describe the high reflectivity of copper (actual energy absorption deviation > 30%) due to the simplification of heat source, and the IMC growth dynamics ignores the interface turbulent effect, with a thickness prediction error of more than 40%, and the multi-physical field full coupling calculation takes several hours and cannot be applied online;

[0006] Data-driven model lacks mechanism fusion: machine learning methods (such as BP neural network) only input static process parameters, ignoring dynamic characteristics such as molten pool oscillation frequency, and small training data size leads to boundary condition failure, and black box characteristics restrict process reverse optimization;

[0007] Therefore, in view of the above problems, a copper-aluminum heterogeneous material laser welding mechanical property prediction model is proposed. SUMMARY

[0008] In order to overcome the above-mentioned defects of the prior art, the embodiments of the present application provide a copper-aluminum heterogeneous material laser welding mechanical property prediction model to solve the problems raised in the background art.

[0009] To achieve the above object, the application provides the following technical scheme: a copper-aluminum heterogeneous material laser welding mechanical property prediction model, comprising:

[0010] The multi-source data acquisition module acquires process parameters, material state parameters and dynamic process parameters in real time, wherein the process parameters include laser power, scanning speed and defocusing amount, the material state parameters include copper layer thickness, aluminum layer thickness and surface oxide film state, and the dynamic process parameters include molten pool oscillation images captured by a high-speed camera system and thermal cycle curves recorded by an infrared temperature measurement system;

[0011] The physical mechanism modeling module is connected to the data acquisition module, a welding process simulation model is constructed through a multi-physical field coupling finite element algorithm, and welding heat affected zone temperature gradient distribution and residual stress field evolution are simulated and calculated based on input process parameters and material state parameters, wherein an adaptive heat source unit dynamically adjusts the energy space distribution mode of a Gaussian heat source according to the difference in copper-aluminum thermal conductivity to compensate for the energy loss caused by the high thermal conductivity of copper, and a stress deformation unit calculates welding deformation and residual stress peak position by coupling a thermal elastic-plastic constitutive equation;

[0012] The data-driven prediction module adopts a double-channel deep neural network architecture, a first channel receives a numerical vector of process parameters and material state parameters, a second channel receives a molten pool geometric feature and thermal cycle feature vector extracted by an image processing unit, and joint prediction values of tensile strength, elongation and interface hardness are output through a feature fusion layer;

[0013] The online decision feedback module compares the predicted tensile strength value with a preset safety threshold in real time, activates a Bayesian optimization engine to generate a process parameter adjustment instruction when the predicted value is lower than the threshold, and completes parameter optimization within 500 ms through a closed-loop control system.

[0014] Preferably, the adaptive heat source unit of the physical mechanism modeling module compensates for the high thermal conductivity of copper by increasing the aluminum side heat flux density, a heat source energy compensation function is defined as: and the copper side compensation coefficient is dynamically increased to 2.4 when the welding direction transitions from aluminum to copper; the stress deformation unit calculates residual stress caused by non-uniform thermal expansion by adopting von Mises yield criterion, and outputs a weld longitudinal stress distribution cloud chart and a maximum stress point position.

[0015] Preferably, the training method of the data-driven prediction module includes a virtual-real data collaborative mechanism, which standardizes and combines simulation data sets covering extreme working conditions of laser power 0.8-3.5kW and scanning speed 1.5-7.0m / min with actual welding experimental data sets; adopts a transfer learning mechanism to pre-train network feature extraction layers in a carbon steel homogenous material welding database, and then fine-tunes full connection layer parameters through copper-aluminum heterogeneous welding data; a dynamic weighted loss function gives a 60% optimization weight to the tensile strength in the back propagation, and gives 30% and 10% weights to the elongation and hardness, respectively.

[0016] Preferably, the quality threshold comparison of the online decision feedback module sets the tensile strength safety threshold to 150MPa, and activates the Bayesian optimizer when the predicted value is lower than 145MPa; the parameter optimization unit adopts an expected improvement (EI) acquisition function to search for optimal solutions in the parameter space of laser power 1.0-3.0kW and scanning speed 2.0-6.0m / min, and outputs the format as a JSON instruction message (including laser power increment ΔP and scanning speed correction amount Δv).

[0017] Preferably, the molten pool image processing unit performs molten pool contour segmentation and dynamic feature extraction, adopts a convolutional neural network with encoding and decoding structure to identify the molten pool boundary in high-speed camera images and eliminate spatter noise, calculates the molten pool length, width-depth ratio, tail inclination angle and oscillation frequency, and constructs a molten pool stability space-time feature vector; wherein the oscillation frequency is extracted through FFT analysis of the coordinates of the molten pool center point in 100 consecutive frames, and when the dominant frequency exceeds 180Hz, it is determined as an unstable molten pool.

[0018] Preferably, the output layer of the data-driven prediction module integrates a confidence assessment unit and a risk warning unit, the confidence assessment generates 100 groups of prediction results and calculates the 95% confidence interval through Monte Carlo Dropout in the inference stage; the risk warning unit triggers an audible and light alarm when the confidence interval width of the tensile strength is greater than the target value 22.5MPa, and marks the high-risk welding parameter combination through OPC UA protocol.

[0019] Preferably, the system deployment architecture includes an edge computing layer and a cloud model management platform, the edge computing layer is integrated in the welding equipment controller, and realizes real-time data acquisition and performance prediction within 50ms delay through a TensorRT acceleration engine; the cloud platform re-trains the model every month by calling historical data, and downloads the SHA-256 checked weight file to the edge device through HTTPS protocol.

[0020] Preferably, the visual interactive interface provides a 3D physical field reconstruction view and parameter sensitivity analysis tools. The 3D view dynamically renders the temperature field (300-1000℃ color scale) and stress field distribution (>200MPa is marked in red); the parameter sensitivity analysis calculates the partial derivative of tensile strength with respect to laser power. Generate a response surface plot and label the highly sensitive regions with gradients > 8 MPa / kW.

[0021] Preferably, the process window is generated through a multi-objective optimization process, and the Pareto optimal solution set is searched by using the non-dominated sorting genetic algorithm (NSGA-II) with the goal of maximizing tensile strength and minimizing the thickness of the intermetallic compound layer. The core process window output is a laser power of 1.8±0.3kW and a scanning speed of 4.0±1.0m / min, and the safe operating range with an intermetallic compound layer thickness of ≤5μm is marked.

[0022] The technical effects and advantages of this invention are as follows:

[0023] Compared to existing technologies, this invention constructs a closed-loop system encompassing multi-source data acquisition, physical mechanism modeling, data-driven prediction, and online decision feedback. First, it utilizes a sensor array to capture multimodal parameters in real time, such as laser power, molten pool oscillation images, and thermal cycling curves. This is combined with an adaptive heat source unit to dynamically compensate for the thermal conductivity differences between copper and aluminum, accurately simulating the temperature field and residual stress evolution. Then, a dual-channel deep neural network is used to fuse process parameter numerical vectors with dynamic molten pool image features. The feature layer jointly outputs predicted values ​​for tensile strength, elongation, and hardness. When the predicted values ​​fall below a safety threshold, a Bayesian optimization engine generates process parameter correction instructions and links the laser actuator, achieving millisecond-level closed-loop control. Simultaneously, a virtual-real data collaborative training mechanism expands the coverage of boundary condition samples, and transfer learning enhances the model's generalization ability. Finally, a stable process window is output through a multi-objective optimization process, reducing experimental dependence while ensuring controlled growth of the interface compound layer and achieving the required mechanical properties. Attached Figure Description

[0024] Fig. 1 This is a system framework diagram of the present invention.

[0025] Fig. 2 This is a flowchart of the dual-channel prediction module of the present invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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 are within the scope of protection of the present invention.

[0027] Example 1

[0028] As attached Figs. 1-2 As shown, the multi-source data acquisition module collects process parameters, material state parameters, and dynamic process parameters in real time. The process parameters include laser power, scanning speed, and defocusing amount. The material state parameters include copper layer thickness, aluminum layer thickness, and surface oxide film state. The dynamic process parameters include molten pool oscillation images captured by the high-speed camera system and thermal cycling curves recorded by the infrared temperature measurement system.

[0029] The physical mechanism modeling module is connected to the data acquisition module. It constructs a welding process simulation model through a multi-physics coupled finite element algorithm. Based on the input process parameters and material state parameters, it simulates and calculates the temperature gradient distribution and residual stress field evolution of the welding heat-affected zone. The adaptive heat source unit dynamically adjusts the energy spatial distribution mode of the Gaussian heat source according to the difference in thermal conductivity between copper and aluminum to compensate for the energy loss caused by the high thermal conductivity of copper. The stress deformation unit coupled with the thermo-elastic-plastic constitutive equation calculates the welding deformation and the peak position of residual stress.

[0030] The data-driven prediction module adopts a dual-channel deep neural network architecture. The first channel receives numerical vectors of process parameters and material state parameters, and the second channel receives geometric features of the molten pool and thermal cycle feature vectors extracted by the image processing unit. The combined predicted values ​​of tensile strength, elongation and interface hardness are output through the feature fusion layer.

[0031] The online decision feedback module compares the predicted tensile strength value with the preset safety threshold in real time. When the predicted value is lower than the threshold, the Bayesian optimization engine is activated to generate process parameter adjustment instructions. The parameters are then optimized within 500ms through the closed-loop control system. The multi-source data acquisition module is implemented through a sensor array integrated into the welding equipment: laser power (0-5kW range) and scanning speed (1-10m / min) are read in real time by the equipment PLC; copper / aluminum layer thickness (accuracy ±0.01mm) is measured at the feed end using a Keyence LJ-V7080 laser thickness gauge; molten pool oscillation images are captured by a Phantom VEO710L high-speed camera (5000fps frame rate, 35° tilt shooting); and thermal cycling curves are recorded by a FLIR A655sc infrared thermal imager (100Hz sampling rate). All data are input into the physical mechanism modeling module after being timestamped according to the IEEE 1588 protocol. This module constructs a three-dimensional thermo-mechanical coupling model based on ANSYS Workbench 2021 R1. The copper and aluminum layers are meshed using SOLID70 elements. The adaptive heat source element dynamically adjusts the energy distribution of the double-ellipsoidal Gaussian heat source according to the copper-aluminum thermal conductivity ratio (65% on the copper side and 35% on the aluminum side). The stress-deformation element couples the thermo-elastic-plastic constitutive equation to output the location of the residual stress peak. The data-driven prediction module uses PyTorch to build a dual-channel network. The first channel inputs a 12-dimensional process / material parameter vector (64 nodes in a fully connected layer), and the second channel inputs 8-dimensional features such as the molten pool aspect ratio and oscillation frequency (128 nodes in an LSTM layer). The feature fusion layer outputs the joint predicted values ​​of tensile strength, elongation, and hardness. The online decision feedback module sets a tensile strength safety threshold of 150 MPa. When the predicted value is lower than 145 MPa, a ΔP / Δv correction command is generated through a Bayesian optimization engine and then transmitted via Modbus. The RTU protocol is transmitted to the laser power controller (response time 100ms) and the motion control card, and closed-loop adjustment is completed within 500ms.

[0032] Example 2

[0033] (1) Heat source and stress calculation

[0034] The adaptive heat source unit of the physical mechanism modeling module compensates for the high thermal conductivity of copper by increasing the heat flux density on the aluminum side. The heat source energy compensation function is defined as: and the compensation coefficient on the copper side is dynamically increased to 2.4 when the welding direction transitions from aluminum to copper. The stress deformation unit uses the von Mises yield criterion to calculate the residual stress caused by non-uniform thermal expansion, and outputs the longitudinal stress distribution cloud map of the weld and the location of the maximum stress point. The adaptive heat source unit defines the energy compensation function in COMSOL. (Initial value of kcu 1.8, initial value of kal 1.2). When the welding direction transitions from aluminum to copper, kcu increases linearly to 2.4 to compensate for high reflection loss. The stress deformation element uses the von Mises yield criterion to calculate the residual stress caused by thermal expansion. Monitoring points are set at intervals of 0.1 mm in the longitudinal direction of the weld to output the location of the maximum stress point (usually 0.2-0.5 mm away from the fusion line) and a three-dimensional stress distribution cloud map. The area >200MPa in the cloud map is marked as a red warning zone.

[0035] (2) Prediction Model Training

[0036] The training method of the data-driven prediction module includes a virtual-real data collaboration mechanism, which standardizes and merges simulation datasets covering extreme working conditions of laser power 0.8-3.5kW and scanning speed 1.5-7.0m / min with actual welding experiment datasets; a transfer learning mechanism is used to first pre-train the network feature extraction layer on a carbon steel homogeneous material welding database, and then fine-tunes the parameters of the fully connected layer using copper-aluminum heterogeneous welding data; a dynamic weighted loss function assigns 60% optimization weight to tensile strength, and 30% and 10% weights to elongation and hardness, respectively, during backpropagation. In the virtual-real data collaboration mechanism, 1000 sets of virtual samples (laser power 0.8-3.5kW) are generated through parametric scanning using ANSYS. The scanning speed was 1.5-7.0 m / min, and the data were merged with 50 sets of actual welding data (Cu110 / AA6061, plate thickness 1.0 mm) after Z-score normalization. The transfer learning mechanism first used the CarbonSteel-Weld public dataset (5000 sets) to pre-train the U-Net encoder and LSTM layer weights. After freezing the feature extraction layer, the fully connected layer was fine-tuned using the copper and aluminum special dataset (200 sets), and the learning rate was set to 0.0001. The dynamic weighted loss function was defined as Loss = 0.6·MSE(σt) + 0.3·MSE(δ) + 0.1·MSE(HV), and the backpropagation used the Adam optimizer for 200 iterations.

[0037] (3) Online decision-making process

[0038] The online decision feedback module sets the tensile strength safety threshold to 150 MPa for quality threshold comparison. When the predicted value is below 145 MPa, the Bayesian optimizer is activated. The parameter optimization unit uses the Expected Improvement (EI) acquisition function to search for the optimal solution in the parameter space of laser power 1.0-3.0 kW and scanning speed 2.0-6.0 m / min. The output format is a JSON command message (containing the laser power increment ΔP and the scanning speed correction Δv). The quality threshold comparison unit presets the tensile strength threshold... The limit is 145MPa (with a 5% margin). When the predicted value σp < 145MPa, the Bayesian optimizer built by the GPyOpt library is activated. The parameter optimization unit uses the expected improvement (EI) as the acquisition function to search for the optimal solution in the space of laser power P∈[1.0,3.0]kW and scanning speed v∈[2.0,6.0]m / min, and outputs JSON format instructions (example: {"ΔP":+0.15,"Δv":-0.3}). The instructions are parsed by the equipment PLC and drive the actuator.

[0039] (4) Molten pool image processing

[0040] The molten pool image processing unit performs molten pool contour segmentation and dynamic feature extraction. It uses a convolutional neural network with an encoder-decoder structure to identify the molten pool boundary in high-speed camera images and eliminate spatter noise. It calculates the molten pool length, aspect ratio, tail tilt angle, and oscillation frequency to construct a spatiotemporal feature vector of molten pool stability. The oscillation frequency is extracted through FFT analysis of the coordinates of the molten pool center point over 100 consecutive frames. When the main frequency exceeds 180Hz, it is determined to be an unstable molten pool. The molten pool contour segmentation uses a U-Net network (input size 256×256, encoder loaded with VGG16 pre-trained weights). After outputting a binary mask, it uses OpenCV to calculate the molten pool length Lp (long side of the maximum circumscribed rectangle), aspect ratio Rwd (width / penetration depth), and tail tilt angle α (angle between the tail tangent and the horizontal plane). The dynamic feature extraction performs FFT analysis on the coordinates of the molten pool center over 100 consecutive frames to extract the main oscillation frequency fosc. When fosc > 180Hz, a molten pool instability flag is sent to the prediction module.

[0041] (5) Confidence level and risk warning

[0042] The output layer of the data-driven prediction module integrates a confidence assessment unit and a risk warning unit. The confidence assessment generates 100 sets of prediction results and calculates a 95% confidence interval during the inference phase using Monte Carlo Dropout. The risk warning unit triggers an audible and visual alarm when the width of the tensile strength confidence interval is greater than the target value of 22.5 MPa, and marks high-risk welding parameter combinations via the OPC UA protocol. Specifically, the confidence assessment unit enables Monte Carlo Dropout with a 30% drop rate during the inference phase, repeats the inference 100 times to generate the tensile strength prediction value distribution, and calculates the 95% confidence interval [μ-1.96σ,μ+1.96σ]. The risk warning unit sets a confidence interval width threshold Wth = 15%·σtarget (Wth = 22.5 MPa when the target strength is 150 MPa). When the actual width Wactual > Wth, it sends an AL101 alarm code to the HMI via the OPC UA protocol and marks high-risk parameter combinations.

[0043] (6) System Deployment Architecture

[0044] The system deployment architecture includes an edge computing layer and a cloud-based model management platform. The edge computing layer is integrated into the welding equipment controller and uses the TensorRT acceleration engine to achieve real-time data acquisition and performance prediction within 50ms latency. The cloud platform retrains the model monthly using historical data and distributes the SHA-256 verified weight files to the edge devices via HTTPS. The edge computing layer is deployed on an NVIDIA Jetson AGX Orin (32GB memory) and integrates the TensorRT acceleration engine to process Modbus TCP data streams (10ms cycle), achieving lightweight segmentation of the molten pool image (latency ≤20ms) and mechanical performance prediction (total latency ≤50ms). The cloud-based model management platform uses an Alibaba Cloud ECS instance (8 cores, 32GB) and retrains the model monthly using historical data warehouse (>100,000 sets). The weight files (.onnx format) are distributed to the edge devices via HTTPS, and version consistency is verified using SHA-256 digests.

[0045] (7) Visual Interaction

[0046] The visual interactive interface provides a 3D physics field reconstruction view and parameter sensitivity analysis tools. The 3D view dynamically renders the temperature field (300-1000℃ color scale) and stress field distribution (>200MPa is marked in red); the parameter sensitivity analysis calculates the partial derivative of tensile strength with respect to laser power. A response surface plot was generated, and highly sensitive regions with gradients > 8 MPa / kW were marked. The 3D physics reconstruction view was developed based on the Unity3D engine, importing .rst result files generated by ANSYS to dynamically render the temperature field (color scale range 300-1000℃) and stress field (highlighted in red for regions > 200 MPa). The parameter sensitivity analysis tool calculated the partial derivative of tensile strength with respect to laser power using the finite difference method. Generate a response surface plot using the Matplotlib library and label highly sensitive regions with gradients > 8 MPa / kW (e.g., gradients reach 10-12 MPa / kW when power > 2.5 kW) with contour lines.

[0047] (8) Process window generation

[0048] A process window is generated through a multi-objective optimization process. The Non-Dominated Sorting Genetic Algorithm (NSGA-II) is used to search for the Pareto optimal solution set with the objectives of maximizing tensile strength and minimizing the intermetallic compound layer thickness. The core output process window is defined as a laser power of 1.8 ± 0.3 kW and a scanning speed of 4.0 ± 1.0 m / min. A safe operating range of ≤ 5 μm intermetallic compound layer thickness is also marked. The multi-objective optimization employs the NSGA-II algorithm from the PyMoo library, with the objective function set as: maximizing tensile strength maxf1(x) = σ. t (x) and minimizing the IMC layer thickness minf2(x)=d IMC (x), (constraints: σt≥160MPa, dIMC≤5μm), decision variable x=(P,v), population size 100, output Pareto optimal solution set after 50 iterations; the core process window is defined as laser power 1.8±0.3kW, scanning speed 4.0±1.0m / min (covering 90% of the optimal solution), and the safe region with interface compound layer thickness ≤5μm is marked as a green area in the interactive interface.

[0049] Example 3: Laser welding of copper and aluminum tabs for new energy battery modules

[0050] 1. Hardware deployment and initial parameter settings

[0051] Integrating sensor arrays at the welding station of the battery module production line:

[0052] Material condition: copper tabs (Cu110, 1.0mm thick) and aluminum tabs (AA6061, 1.0mm thick). Thickness fluctuations (±0.02mm) were monitored in real time using a Keyence LJ-V7080 laser thickness gauge.

[0053] Process parameters: initial laser power 1.6kW, scanning speed 5.0m / min, defocusing amount +1mm (controlled by Siemens S7-1500 PLC);

[0054] Dynamic monitoring: The Phantom VEO710L high-speed camera (5000fps, 35° tilt) captures the molten pool oscillation, and the FLIRA655sc infrared thermal imager records the thermal cycling curve. The data is aligned according to the IEEE 1588 protocol.

[0055] 2. Anomaly Detection and Real-time Prediction

[0056] During the welding process, the data-driven prediction module detected that the molten pool oscillation frequency reached 210Hz (>180Hz threshold), and the dual-channel network output predicted values:

[0057] Tensile strength: 142 MPa (below the safety threshold of 145 MPa);

[0058] Elongation: 4.8%;

[0059] Confidence interval: [132, 152] MPa (width 20 MPa > 22.5 MPa threshold);

[0060] The risk warning unit triggers the AL101 audible and visual alarm, marking the current parameters as a high-risk condition.

[0061] 3. Bayesian online optimization

[0062] The online decision feedback module activates the Bayesian optimization engine (GPyOpt library):

[0063] Acquisition function: Desired improvement (EI);

[0064] Parameter space: P∈[1.0,3.0]kW, v∈[2.0,6.0]m / min;

[0065] Output command: {"ΔP":+0.25,"Δv":-0.8} (i.e., power increases to 1.85kW, speed decreases to 4.2m / min).

[0066] The commands are transmitted via Modbus RTU to the laser controller (100ms response) and motion control card, and the parameter adjustment is completed within 500ms.

[0067] 4. Performance verification after optimization

[0068] Physical mechanism model simulation: The adaptive heat source unit increases the copper side compensation coefficient to 2.1, and the peak residual stress decreases from 218MPa to 185MPa;

[0069] Prediction module output: Tensile strength increased to 165MPa (confidence interval [160,170]MPa), molten pool oscillation frequency decreased to 155Hz;

[0070] Actual welding test (GB / T 2651-2008 standard):

[0071] (1) Tensile strength: 163-171 MPa (fluctuation ±2.4%);

[0072] (2) IMC layer thickness: 4.3 μm (metallographic measurement);

[0073] (3) The weld is free of cracks and porosity defects (X-ray inspection pass rate is 100%).

[0074] 5. Process window output

[0075] The multi-objective optimization module (NSGA-II algorithm) generates the core process window based on 200 sets of historical data:

[0076] Laser power: 1.8±0.3kW (optimal value 1.85kW);

[0077] Scanning speed: 4.0 ± 1.0 m / min (optimal value 4.2 m / min);

[0078] Safe operating area: IMC layer thickness ≤ 5μm (actual value 4.3μm, marked as green area).

[0079] Finally, the following points should be noted: First, in the description of this application, it should be noted that, unless otherwise specified and limited, the terms "installation", "connection", and "linkage" should be interpreted broadly, and can be mechanical or electrical connections, or internal connections between two components, or direct connections. "Up", "down", "left", "right", etc. are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may change.

[0080] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other.

[0081] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A model for predicting the mechanical properties of laser welding copper-aluminum dissimilar materials, characterized in that, include: The multi-source data acquisition module collects process parameters, material state parameters, and dynamic process parameters in real time. The process parameters include laser power, scanning speed, and defocusing amount. The material state parameters include copper layer thickness, aluminum layer thickness, and surface oxide film state. The dynamic process parameters include molten pool oscillation images captured by the high-speed camera system and thermal cycling curves recorded by the infrared temperature measurement system. The physical mechanism modeling module is connected to the data acquisition module. It constructs a welding process simulation model through a multi-physics coupled finite element algorithm. Based on the input process parameters and material state parameters, it simulates and calculates the temperature gradient distribution and residual stress field evolution of the welding heat-affected zone. The adaptive heat source unit dynamically adjusts the energy spatial distribution mode of the Gaussian heat source according to the difference in thermal conductivity between copper and aluminum to compensate for the energy loss caused by the high thermal conductivity of copper. The stress deformation unit coupled with the thermo-elastic-plastic constitutive equation calculates the welding deformation and the peak position of residual stress. The data-driven prediction module adopts a dual-channel deep neural network architecture. The first channel receives numerical vectors of process parameters and material state parameters, and the second channel receives geometric features of the molten pool and thermal cycle feature vectors extracted by the image processing unit. The combined predicted values ​​of tensile strength, elongation and interface hardness are output through the feature fusion layer. The online decision feedback module compares the predicted tensile strength value with the preset safety threshold in real time. When the predicted value is lower than the threshold, the Bayesian optimization engine is activated to generate process parameter adjustment instructions, and the parameter optimization is completed within 500ms through the closed-loop control system.

2. The mechanical property prediction model for laser welding of copper-aluminum heteromaterials according to claim 1, characterized in that: The adaptive heat source unit of the physical mechanism modeling module compensates for the high thermal conductivity of copper by increasing the heat flux density on the aluminum side. The heat source energy compensation function is defined as: and the compensation coefficient on the copper side is dynamically increased to 2.4 when the welding direction transitions from aluminum to copper. The stress deformation unit uses the von Mises yield criterion to calculate the residual stress caused by non-uniform thermal expansion and outputs the longitudinal stress distribution cloud map of the weld and the location of the maximum stress point.

3. The mechanical property prediction model for laser welding of copper-aluminum heteromaterials according to claim 1, characterized in that: The training method of the data-driven prediction module includes a virtual-real data collaboration mechanism, which standardizes and merges the simulation dataset covering extreme working conditions of laser power 0.8-3.5kW and scanning speed 1.5-7.0m / min with the actual welding experiment dataset; a transfer learning mechanism is used to first pre-train the network feature extraction layer on the carbon steel homogeneous material welding database, and then fine-tunes the parameters of the fully connected layer through copper-aluminum heterogeneous welding data; the dynamic weighted loss function assigns 60% optimization weight to tensile strength, and 30% and 10% weights to elongation and hardness, respectively, during backpropagation.

4. The mechanical property prediction model for laser welding of copper-aluminum heteromaterials according to claim 1, characterized in that: The online decision feedback module sets the tensile strength safety threshold to 150MPa for quality threshold comparison. When the predicted value is lower than 145MPa, the Bayesian optimizer is activated. The parameter optimization unit uses the expectation improvement (EI) acquisition function to search for the optimal solution in the parameter space of laser power 1.0-3.0kW and scanning speed 2.0-6.0m / min. The output format is a JSON command message (including laser power increment ΔP and scanning speed correction Δv).

5. The mechanical property prediction model for laser welding of copper-aluminum heteromaterials according to claim 1, characterized in that: The molten pool image processing unit performs molten pool contour segmentation and dynamic feature extraction. It uses a convolutional neural network with an encoding and decoding structure to identify the molten pool boundary in high-speed camera images and eliminate spatter noise. It calculates the molten pool length, aspect ratio, tail tilt angle and oscillation frequency, and constructs a spatiotemporal feature vector of molten pool stability. The oscillation frequency is extracted by FFT analysis of the center point coordinates of the molten pool for 100 consecutive frames. When the main frequency exceeds 180Hz, it is determined to be an unstable molten pool.

6. The mechanical property prediction model for laser welding of copper-aluminum heteromaterials according to claim 1, characterized in that: The output layer of the data-driven prediction module integrates a confidence assessment unit and a risk warning unit. The confidence assessment uses Monte Carlo Dropout to generate 100 sets of prediction results and calculate a 95% confidence interval during the inference stage. The risk warning unit triggers an audible and visual alarm when the width of the tensile strength confidence interval is greater than the target value of 22.5 MPa, and marks high-risk welding parameter combinations through the OPC UA protocol.

7. The mechanical property prediction model for laser welding of copper-aluminum heteromaterials according to claim 1, characterized in that: The system deployment architecture includes an edge computing layer and a cloud model management platform. The edge computing layer is integrated into the welding equipment controller and uses the TensorRT acceleration engine to achieve real-time data acquisition and performance prediction within 50ms latency. The cloud platform retrains the model monthly using historical data and distributes the SHA-256 verified weight file to edge devices via HTTPS.

8. The mechanical property prediction model for laser welding of copper-aluminum heteromaterials according to claim 1, characterized in that: The visual interactive interface provides a 3D physics field reconstruction view and parameter sensitivity analysis tools. The 3D view dynamically renders the temperature field (300-1000℃ color scale) and stress field distribution (>200MPa is marked in red); the parameter sensitivity analysis calculates the partial derivative of tensile strength with respect to laser power. Generate a response surface plot and label the highly sensitive regions with gradients > 8 MPa / kW.

9. The mechanical property prediction model for laser welding of copper-aluminum heteromaterials according to claim 1, characterized in that: The process window is generated by multi-objective optimization process. The non-dominated sorting genetic algorithm (NSGA-II) is used to search for the Pareto optimal solution set with the goal of maximizing tensile strength and minimizing the thickness of the intermetallic compound layer. The core process window output is a laser power of 1.8±0.3kW and a scanning speed of 4.0±1.0m / min. The safe operating range with an intermetallic compound layer thickness of ≤5μm is marked.

Images