Laser welding defect identification method for copper-aluminum heterogeneous material

Abstract

The invention relates to the field of laser welding quality detection, in particular to a multi-modal fusion identification method and system for laser welding defects of copper-aluminum heterogeneous materials. Molten pool images, temperature fields and acoustic emission signals are synchronously collected through a visible light high-speed camera, a thermal infrared imager and an acoustic emission sensor which are coaxially integrated; the width oscillation amplitude, the trailing edge solidification angle, the interface temperature gradient, the acoustic emission characteristic frequency band energy ratio and the burst signal counting rate of the molten pool are extracted; constructing a fusion vector input adaptive weighted deep neural network model, and identifying four types of defects of pores, cracks, incomplete fusion and interface brittle compound layer standard exceeding in real time; and the identification result triggers laser power modulation or protective gas proportion adjustment. The problems that in the prior art, multi-physical field perception is insufficient, the feature adaptability of heterogeneous materials is poor, and real-time control is lack are solved, the defect detection rate reaches 98.2%, and control response delay lt is achieved; and 50ms.

Description

Technical Field

[0001] This invention relates to the field of laser welding quality inspection, and more specifically, to a method for identifying defects in laser welding of copper-aluminum dissimilar materials. Background Technology

[0002] Laser welding of copper-aluminum dissimilar materials is widely used in new energy batteries, power electronics, and other fields. However, due to the significant differences in the physicochemical properties of the two materials (copper has a melting point of 1083℃ and a thermal conductivity of 401 W / m·K, while aluminum has a melting point of 660℃ and a thermal conductivity of 237 W / m·K), defects such as porosity, cracks, lack of fusion, and brittle intermetallic compound layers are easily generated during the welding process. Traditional manual quality inspection and single-sensor detection methods are insufficient to meet the requirements for high-precision online identification. There is an urgent need to develop multimodal collaborative intelligent defect identification technology. Existing technologies still have certain limitations. 1. Insufficient multi-physics sensing capability: Existing methods mostly rely on a single sensing mode (such as visible light vision or infrared thermometry), which cannot simultaneously capture multi-dimensional physical information of the welding process. Visible light imaging is easily affected by plasma interference, leading to distortion of the molten pool profile; infrared thermometry has difficulty distinguishing the difference in thermal radiation at the copper-aluminum interface; and acoustic emission monitoring ignores the dynamic behavior of the molten pool. This single-mode detection results in a false negative rate of over 30% for porosity and microcracks, especially lacking effective monitoring methods for brittle compound layers at the interface.

[0003] 2. Poor adaptability to heterogeneous material characteristics: Traditional defect identification algorithms are directly adapted from welding of homogeneous materials, failing to consider the unique defect characteristics caused by the differences in the thermophysical properties of copper and aluminum. For example, they neglect the sensitivity of the solidification angle along the back edge of the molten pool to incomplete fusion; they fail to establish a correlation model between the interface temperature gradient and the formation of brittle compounds; and they lack targeted analysis of the frequency band characteristics of acoustic emission signals from copper and aluminum. This results in the algorithm's accuracy in identifying interface defects generally being below 65%.

[0004] 3. Lack of Real-Time Closed-Loop Control: Current systems are mostly limited to offline detection and have not achieved a "recognition-control" closed loop. Defect identification results cannot be fed back to welding process parameter adjustments (such as dynamic modulation of laser power and optimization of shielding gas composition) in real time, resulting in a disconnect between identification and process optimization. Experiments show that 70% of defects found after welding can be avoided through real-time intervention, but existing technologies cannot complete the decision-making response within a 50ms time window.

[0005] Therefore, a method for identifying defects in laser welding of copper-aluminum dissimilar materials is proposed to address the above problems. Summary of the Invention

[0006] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a method for identifying defects in laser welding of copper-aluminum dissimilar materials, thereby addressing the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for identifying defects in laser welding of copper-aluminum dissimilar materials, comprising the following steps: S1. Synchronously acquire welding process data through a coaxial integrated multi-sensor system: use a visible light high-speed camera with an 808nm bandpass filter to capture dynamic image sequences of the molten pool, use an infrared thermal imager with a thermal sensitivity ≤5mK to record time-series data of temperature field distribution, and use a waveguide rod coupled acoustic emission sensor with a resonant frequency of 150kHz to acquire acoustic emission time-domain signals. S2. Spatiotemporal registration and feature extraction of multimodal data: Perform molten pool contour segmentation based on attention mechanism on visible light images to extract molten pool width oscillation amplitude and trailing solidification angle; identify the boundary of copper-aluminum heat-affected zone and calculate the interface temperature gradient on infrared thermal images; perform wavelet packet decomposition on acoustic emission signals and extract the energy proportion of specific frequency bands and burst signal count rate. S3. Construct a defect-sensitive fusion vector from the features extracted in step (S2) and input it into a pre-trained adaptive weighted deep neural network model; S4. Based on the probability distribution of the model output, four types of defects are identified in real time: porosity, cracks, lack of fusion, and excessive interface brittle compound layer.

[0008] Preferably, the sensing system configuration in step (S1) includes: a visible light high-speed camera acquiring data at a rate of not less than 5000 frames / second, and using an 808nm filter to suppress plasma radiation interference; an infrared thermal imager with a spatial resolution higher than 0.1 mm / pixel, and clock synchronization with the visible light camera achieved through hardware triggering; an acoustic emission sensor rigidly coupled to a range of 10 mm from the weld seam via a waveguide rod; and the sampling time deviation of the three sensors controlled within 1 microsecond.

[0009] Preferably, the molten pool contour segmentation in step (S2) specifically includes: adopting a U-Net network architecture with an embedded spatial attention module, whose encoder extracts multi-scale features based on the ResNet34 backbone network, integrates prior knowledge of the difference in reflectivity between copper and aluminum materials in the decoding stage, corrects the segmentation results through threshold constraints, and finally outputs the temporal sequence of the molten pool width oscillation amplitude ΔW and the trailing solidification angle θ.

[0010] Preferably, the interface temperature gradient calculation in step (S2) includes: accurately calibrating the coordinate set of the copper-aluminum heteromaterial interface in the infrared thermal image, extending a region of 0.5 mm on each side along the interface normal as the temperature gradient calculation area, and calculating the temperature change per unit distance at a frequency of 100 Hz as the standardized gradient value ∇T.

[0011] Preferably, the adaptive weighted deep neural network model in step (S3) includes a feature weighting layer, a multi-branch feature fusion module, and a classification output layer. The feature weighting layer dynamically assigns weight coefficients based on the contribution of features to defect identification. The weight adjustment is based on the joint optimization of the loss function gradient and feature information entropy. The multi-branch feature fusion module uses a one-dimensional convolutional neural network to process acoustic emission features, a long short-term memory network to analyze temperature time-series features, and a three-dimensional convolutional neural network to extract spatiotemporal features of the molten pool image. The classification output layer generates probability values ​​for four types of defects using the Softmax function.

[0012] Preferably, the model training adopts a cross-material transfer learning strategy, pre-trains the network infrastructure using a stainless steel welding dataset, and fine-tunes the fully connected layer parameters on a copper-aluminum heterogeneous material-specific dataset. The loss function integrates Focal Loss and distribution alignment metric to balance class imbalance and enhance domain adaptability.

[0013] Preferably, the method also includes real-time closed-loop control, which triggers a pulse modulation control strategy for laser power when the probability of porosity defects exceeds 80%, automatically adjusts the mixing ratio of helium and argon when the probability of the interface brittle compound layer exceeding the standard exceeds 75%, and the total delay time from defect identification to execution of the control command is less than 50 milliseconds.

[0014] Preferably, the acoustic emission feature extraction in step (S2) includes: performing 5-layer wavelet packet decomposition using the Daubechies wavelet basis, and calculating the energy ratio K of the two characteristic frequency bands 62.5-125kHz and 125-250kHz. When the energy ratio K is greater than 2.5 and the burst signal count rate exceeds 200 times / second, a high-risk crack warning is activated.

[0015] Preferably, a knowledge base for copper-aluminum interface reaction is constructed: thermodynamic generation parameters of intermetallic compounds in copper-aluminum alloys are pre-stored, and the degree of interface reaction is inverted based on real-time temperature field data. When the inverted value exceeds a preset threshold, the identification priority of brittle compound layer defects at the interface is increased.

[0016] A laser welding quality monitoring system, comprising: Multimodal sensing unit: a visible light camera, an infrared thermal imager, and an acoustic emission sensor integrated with an explosion-proof housing; Real-time processing unit: An embedded industrial computer equipped with a graphics processor, used to execute feature extraction and defect identification algorithms; Human-computer interaction unit: A visual interface that displays the dynamics of the molten pool, the temperature field distribution, and the probability of defects; Control and execution unit: Feedback controller connected to the laser generator and gas supply device.

[0017] The technical effects and advantages of this invention are as follows: 1. Multi-physics holographic sensing: By coordinating visible light, infrared, and acoustic emission sensing modes, this approach overcomes the limitations of single-sensor systems. High-speed visible light imaging combined with an 808nm filter suppresses plasma interference, achieving sub-pixel-level segmentation of the molten pool contour; high-precision infrared thermal imager analyzes the temperature gradient at the copper-aluminum interface; and acoustic emission sensors accurately capture crack propagation signals within the material. Experiments show that this approach increases the detection rate of porosity and microcracks to 98.2%, while reducing the false negative rate to below 1.8% (a 68% improvement over existing technologies).

[0018] 2. Heterogeneous Material-Specific Feature Engine: A feature extraction algorithm was developed specifically for the differences in the physical properties of copper and aluminum: ① A melt pool segmentation network based on an attention mechanism is used to integrate material reflection priors to accurately extract features such as the solidification angle; ② A mapping model between the interface temperature gradient and the thickness of the brittle compound is established; ③ Wavelet packet decomposition focuses on the characteristic frequency bands of copper and aluminum in the 50-200kHz range. In the testing of new energy battery connectors, the accuracy rate of interface defect identification reached 93.5% (28.5% improvement compared to traditional methods).

[0019] 3. Millisecond-level real-time closed-loop control: Constructing a full-link system of "identification-decision-execution": ① Triggering laser power pulse modulation (frequency 1-5kHz) when the porosity probability >80%; ② Switching the He / Ar mixed gas ratio when the compound layer exceedance probability >75% (response delay <50ms). Dynamic control of the welding process reduces the incidence of avoidable defects to 8.3% (achieving an 89% suppression rate compared to the existing 70% potential avoidable defects).

[0020] 4. Cross-domain adaptive intelligent decision-making: Employing transfer learning and adaptive weighting mechanisms: ① Pre-training the network foundation using stainless steel welding big data; ② Aligning copper and aluminum data distributions through an MMD domain adaptive loss function; ③ The dynamic weight layer automatically adjusts the fusion coefficients based on feature contribution. This method maintains a 91.7% recognition stability in 200 sets of unknown process parameter samples (a 42% improvement over the direct transfer model). Attached Figure Description

[0021] Figure 1 This is a system framework diagram of the present invention. Detailed Implementation

[0022] 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.

[0023] As attached Figure 1As shown, (1) a method for identifying defects in laser welding of copper-aluminum dissimilar materials includes the following steps: S1. Synchronously acquire welding process data through a coaxial integrated multi-sensor system: use a visible light high-speed camera with an 808nm bandpass filter to capture dynamic image sequences of the molten pool, use an infrared thermal imager with a thermal sensitivity ≤5mK to record time-series data of temperature field distribution, and use a waveguide rod coupled acoustic emission sensor with a resonant frequency of 150kHz to acquire acoustic emission time-domain signals. S2. Spatiotemporal registration and feature extraction of multimodal data: Perform molten pool contour segmentation based on attention mechanism on visible light images to extract molten pool width oscillation amplitude and trailing solidification angle; identify the boundary of copper-aluminum heat-affected zone and calculate the interface temperature gradient on infrared thermal images; perform wavelet packet decomposition on acoustic emission signals and extract the energy proportion of specific frequency bands and burst signal count rate. S3. Construct a defect-sensitive fusion vector from the features extracted in step (S2) and input it into a pre-trained adaptive weighted deep neural network model; S4. Based on the probability distribution of the model output, four types of defects are identified in real time: porosity, cracks, lack of fusion, and excessive interfacial brittle compound layer. Specifically, in the welding of copper-aluminum electrodes for new energy batteries, an IPG YLS-6000 fiber laser (3.5kW power, 0.2mm spot diameter) is used, along with a coaxially integrated Basler acA2000-510km high-speed camera (5000fps), a FLIR A6751sc infrared thermal imager (2.5mK thermal sensitivity), and Physical Acoustics. Data was synchronously acquired using a PICO acoustic emission sensor (resonant frequency 150kHz); the molten pool contour was segmented using an improved U-Net algorithm (the encoder used ResNet34 pre-trained weights), and the molten pool width oscillation amplitude ΔW = 0.12-0.25mm and the trailing solidification angle θ = 45-65° were extracted; the interface temperature gradient ∇T = ΔT / 1mm was calculated based on the infrared image (ΔT is the temperature difference within a 0.5mm normal region); the acoustic emission signal was decomposed into Db6 wavelet packets at 5 levels to extract the energy percentage E% in the 125-250kHz frequency band and the burst signal count rate R > 80dB; a feature vector V = [ΔW, θ, ∇T, E%, R] was constructed and input into an adaptive weighted DNN model, and a control command was triggered when the output porosity probability > 80%.

[0024] (2) The sensor system configuration in step (S1) includes: a visible light high-speed camera acquiring data at a rate of not less than 5000 frames / second, and using an 808nm filter to suppress plasma radiation interference; an infrared thermal imager with a spatial resolution higher than 0.1 mm / pixel, and clock synchronization with the visible light camera via hardware triggering; an acoustic emission sensor rigidly coupled to a distance of 10 mm from the weld seam via a waveguide rod; and the sampling time deviation of the three sensors controlled within 1 microsecond. The visible light camera integrates an 808nm bandpass filter (2mm thick) via an M72 threaded interface and is installed at a 30° angle to the laser beam; the infrared thermal imager is spatially calibrated (accuracy ±0.05mm) via a copper-aluminum dual-material calibration plate, and synchronized with the camera via an FPGA to generate a trigger pulse (rising edge ≤10ns); the acoustic emission sensor is rigidly coupled to a distance of 8 mm from the weld seam via a zirconia waveguide rod (φ3mm×50mm); and the three sensors are connected to an NI PXIe-5170R acquisition card (sampling rate 1GS / s) via a coaxial cable. Clock synchronization adopts IEEE 1588v2 protocol (deviation ≤ 800ns).

[0025] (3) The specific steps of the molten pool contour segmentation in step (S2) include: adopting a U-Net network architecture with an embedded spatial attention module, the encoder extracts multi-scale features based on the ResNet34 backbone network, integrates prior knowledge of the difference in reflectivity between copper and aluminum materials in the decoding stage, corrects the segmentation results through threshold constraints, and finally outputs the temporal sequence of the molten pool width oscillation amplitude ΔW and the trailing solidification angle θ. The molten pool segmentation is implemented as follows: the improved U-Net model (input size 512×512) is loaded on the NVIDIA Jetson AGX Orin embedded system, the encoder uses ResNet34 to extract 5 levels of feature maps (maximum downsampling rate 1 / 32), the decoder embeds a spatial attention module (calculates the attention weight α=softmax(Conv(F)) of feature map F), introduces copper and aluminum reflectivity difference constraints in the skip connection layer (copper reflectivity threshold 0.6, aluminum threshold 0.3), and finally outputs a binary mask and calculates the trailing solidification angle θ of the molten pool.

[0026] (4) The interface temperature gradient calculation in step (S2) includes: accurately calibrating the coordinate set of the copper-aluminum heterostructure interface in the infrared thermal image, extending a region of 0.5 mm on each side along the interface normal as the temperature gradient calculation area, and calculating the temperature change per unit distance at a frequency of 100 Hz as the standardized gradient value ∇T. The interface temperature gradient calculation is as follows: first, locate the alumina marker point in the infrared image (pre-set at the copper-aluminum interface, diameter 0.1 mm), and generate a 1 mm wide analysis band (resolution 0.02 mm / pixel) along the normal with the marker point as the center; use bicubic interpolation to calculate the temperature distribution curve T(x) in the analysis band, and calculate the gradient value by the central difference method: ∇T=|T(x+0.5)-T(x-0.5)| / 1 mm (unit ℃ / mm), and output the gradient sequence once every 100 ms (typical value range 50-120 ℃ / mm).

[0027] (5) The adaptive weighted deep neural network model in step (S3) includes a feature weighting layer, a multi-branch feature fusion module, and a classification output layer. The feature weighting layer dynamically assigns weight coefficients based on the contribution of features to defect identification. The weight adjustment is based on the joint optimization of the loss function gradient and feature information entropy. The multi-branch feature fusion module uses a one-dimensional convolutional neural network to process acoustic emission features, a long short-term memory network to analyze temperature time-series features, and a three-dimensional convolutional neural network to extract spatiotemporal features of the molten pool image. The classification output layer generates probability values ​​for four types of defects using the Softmax function. The adaptive weighted deep neural network... N implementation: The model input layer receives a 5-dimensional feature vector, and the weighted layer is set with a trainable parameter matrix W∈R^(5×5) (initial value identity matrix); in the multi-branch fusion module, the acoustic emission features are processed by 1D-CNN (convolution kernel [1×5], stride 2), the temperature features are input into a bidirectional LSTM (128 hidden units), and the molten pool image sequence is input into a 3D-CNN (kernel size [3×3×3]); the fused features are fully connected to a 512-neuron layer, and the output layer Softmax generates four types of defect probabilities; training uses the Adam optimizer (lr=0.001), and the dynamic weights are updated through gradient backpropagation.

[0028] (6) The model training adopts a cross-material transfer learning strategy. The network infrastructure is pre-trained using the stainless steel welding dataset, and the parameters of the fully connected layers are fine-tuned on the copper-aluminum heterogeneous material dataset. The loss function combines Focal Loss and distribution alignment metric to balance class imbalance and enhance domain adaptability. The transfer learning is implemented as follows: In the pre-training stage, the 304 stainless steel pulsed laser welding dataset (20,000 samples) is used to train the basic network (ResNet34 backbone) in the PyTorch framework; in the fine-tuning stage, the copper-aluminum heterogeneous welding dataset (1,500 samples) is loaded, the encoder parameters are frozen, and only the fully connected layers are trained; the loss function is jointly optimized by Focal Loss (γ=2.0) and MMD distribution alignment (λ1=0.7, λ2=0.3), and the training cycle is 50 rounds (batch_size=32).

[0029] (7) The method also includes real-time closed-loop control. When the probability of porosity defects exceeds 80%, the pulse modulation control strategy of laser power is triggered. When the probability of the interface brittle compound layer exceeds 75%, the mixing ratio of helium and argon is automatically adjusted. The total delay time from defect identification to execution of the control command is less than 50 milliseconds. The closed-loop control execution is as follows: when the probability of porosity is >80%, the power modulation command is sent to the laser via the EtherCAT protocol (waveform: square wave, frequency 2kHz, duty cycle 60-80% dynamically adjusted); when the probability of compound layer is >75%, the SMC proportional valve is controlled to adjust the He / Ar mixing ratio (He ratio increases from 30% to 70%), and the gas flow meter provides real-time feedback of the actual ratio (control accuracy ±3%). The execution delay test is as follows: the time from probability output to valve action is 23±5ms (based on oscilloscope measurement of IO signal).

[0030] (8) The acoustic emission feature extraction in step (S2) includes: using the Daubechies wavelet basis to perform 5-layer wavelet packet decomposition, and calculating the energy ratio K of the two characteristic frequency bands 62.5-125kHz and 125-250kHz. When the energy ratio K is greater than 2.5 and the burst signal count rate exceeds 200 times / second, the high-risk crack warning is activated. The acoustic emission warning is implemented as follows: the original signal (sampling rate 1MHz) is decomposed into 5 layers of Db6 wavelet packets, and the signals of the 4th sub-band (62.5-125kHz) and the 5th sub-band (125-250kHz) are reconstructed; the energy ratio K is calculated as K=∫|S4(t)|²dt / ∫|S5(t)|²dt; the burst signal count adopts the short-time energy over-threshold method (threshold = 5 times the root mean square, time window 1ms); when K>2.5 and R>200 times within the 10ms window, a level 3 sound and light alarm is triggered (industrial PC buzzer frequency 2kHz).

[0031] (9) Constructing a knowledge base for copper-aluminum intermetallic reactions: Pre-store the thermodynamic generation parameters of intermetallic compounds in copper-aluminum alloys, and invert the intermetallic reaction degree value based on real-time temperature field data. When the inverted value exceeds the preset threshold, the identification priority of brittle compound layer defects at the interface is increased. The implementation of the intermetallic reaction knowledge base is as follows: Store the IMC thermodynamic parameters such as CuAl2 (ΔH=-36.8kJ / mol) and Cu9Al4 (ΔH=-28.7kJ / mol) in the SQLite database; acquire the temperature T(x,y,t) of the interface coordinate point in the infrared image in real time, and calculate the reaction degree η=Σexp(-ΔH / (R_g·T))·Δt (R_g=8.314 J / mol·K) through the Arrhenius equation; when η>3.7 (calibration threshold), increase the weight of ∇T by 20% in the feature weighting layer.

[0032] (10) Laser welding quality monitoring systems, including: Multimodal sensing unit: a visible light camera, an infrared thermal imager, and an acoustic emission sensor integrated with an explosion-proof housing; Real-time processing unit: An embedded industrial computer equipped with a graphics processor, used to execute feature extraction and defect identification algorithms; Human-computer interaction unit: A visual interface that displays the dynamics of the molten pool, the temperature field distribution, and the probability of defects; Control and execution unit: Feedback controller connecting the laser generator and gas supply device. System hardware implementation: The sensing unit adopts a stainless steel explosion-proof housing (IP67) and has a built-in thermoelectric cooler to maintain an operating temperature of 35°C; the processing unit is equipped with Jetson AGX Orin (32GB RAM, CUDA core 1792); the human-machine interface is equipped with a 12-inch industrial touch screen (resolution 1920×1080), which displays the dynamic outline of the molten pool (red), temperature pseudo-color map (blue-yellow-red level), and defect probability radar map in real time; the control unit is connected to the laser (IPG YLS controller) and mass flow meter (Alicat Scientific PC series) through the Profinet interface. Example

[0033] Phase 1: System Initialization and Parameter Configuration 1. Hardware Deployment The coaxial integrated probe (including a visible light camera, an infrared thermal imager, and an acoustic emission sensor) is fixed to the side of the welding head, with the probe axis forming a 30° angle with the laser beam. The visible light camera is equipped with an 808nm bandpass filter, set to a resolution of 1280×1024@5000fps, and the white balance mode is set to "metal welding". The infrared thermal imager is set to measure temperatures from 300 to 1500℃, with a spatial calibration of 0.08 mm / pixel. Emissivity parameters are: copper 0.15 / aluminum 0.25. The acoustic emission sensor is coupled to a location 8 mm from the weld via a waveguide rod, with a gain set to 60 dB and a sampling frequency of 1 MHz. 2. Software Configuration # Load the pre-trained model in the real-time processing unit model = load_model("defect_detection_v3.h5") # Set feature extraction parameters config = { "Molten pool segmentation threshold": 0.7, # Based on the difference in reflectivity between copper and aluminum Temperature gradient calculation step size: 0.5, # Unit: mm Wavelet packet decomposition level: 5, "Warning threshold": {"Stomata": 0.8, "Compound layer": 0.75}} Phase Two: Synchronous Monitoring of the Welding Process 1. Synchronous acquisition of multi-source data When welding starts, the three sensors are synchronously acquired via FPGA hardware (clock deviation ≤ 1μs): Visible light captures dynamic sequences of the molten pool; Infrared recording of temperature field distribution; Acoustic emission acquisition of time-domain signals; Data is transmitted to the GPU processing unit (NVIDIA Jetson AGX Orin) in real time. 2. Spatiotemporal registration method Establish a three-dimensional spatial mapping with the laser irradiation point as the origin:

[0034] The acoustic emission signal is time-scaled using wavelet transform (time-domain offset compensation Δt=0.3ms). Phase 3: Defect Feature Extraction 1. Geometric Analysis of the Molten Pool Execute the improved U-Net segmentation algorithm: Input: Visible light image sequence → Encoder (ResNet34 extracts 5 levels of features) Decoder embeds spatial attention module, weight calculation: Output: Binary mask of molten pool profile → Calculate width oscillation amplitude ΔW = 0.15 ± 0.03 mm 2. Thermodynamic Feature Extraction Infrared image processing workflow: Thermal image preprocessing → Copper-aluminum interface identification → Normal extension of 0.5mm region → Calculation of gradient ∇T = ΔT / 1mm Typical values: Copper side ∇T_max = 85℃ / mm, Aluminum side ∇T_max = 62℃ / mm 3. Analysis of Acoustic Emission Characteristics Wavelet packet decomposition (Db6 basis functions): subband layer Frequency band (kHz) Energy percentage 4 62.5-125 42.3% 5 125-250 18.7% The characteristic ratio K is calculated to be 42.3 / 18.7 = 2.26 (which is lower than the crack threshold of 2.5). Phase 4: Intelligent Decision-Making and Real-Time Control 1. Defect Identification Reasoning Construct the feature vector V = [ΔW=0.18, θ=52°, ∇T=73, E%=42.3, R=185] Input adaptive weighted DNN model: The output weight matrix of the feature-weighted layer is: W=[0.32, 0.21, 0.28, 0.12, 0.07] Multi-branch fusion: 1D-CNN processes acoustic features → LSTM analyzes temperature time series → 3D-CNN extracts spatiotemporal features from images Softmax output probabilities: P(pores) = 65%, P(compound layer) = 82% 2. Closed-loop control execution When the probability of compound layer is 82% > threshold 75%: Send the command to the gas controller: "He:Ar=3:1" (original ratio 1:1) PID control valve response time: 23ms (meets <50ms requirement) Verification of control effect: The thickness of the interface compound layer decreased from 8.2 μm to 3.7 μm. Phase 5: System Verification and Optimization 1. Performance Testing Defect types Sample size Detection rate False alarm rate pores 120 98.3% 1.2% crack 85 97.6% 0.9% compound layer 150 93.5% 2.1% 2. Online model updates Automatically initiate transfer learning after every 50 welding operations are completed: if sample_count % 50 == 0: fine_tune_model(learning_rate=1e-4, epochs=10) # Incrementally update the fully connected layer 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. 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. 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 method for identifying defects in laser welding of copper-aluminum heterogeneous materials, characterized by, The method comprises the following steps: S1, synchronously collecting welding process data by a coaxial integrated multi-sensor system: a visible light high-speed camera with an 808 nm band-pass filter is used to capture a sequence of dynamic images of a molten pool, an infrared thermal imager with a thermal sensitivity of ≤5 mK is used to record time-series data of temperature field distribution, and a waveguide rod coupled resonant frequency 150 kHz acoustic emission sensor is used to obtain acoustic emission time domain signals; S2, time-space registration and feature extraction of multi-modal data: performing molten pool contour segmentation based on an attention mechanism on visible light images to extract molten pool width oscillation amplitude and rear solidification angle, identifying the copper-aluminum heat affected zone boundary on infrared thermal images and calculating the interface temperature gradient, and performing wavelet packet decomposition on acoustic emission signals and extracting specific frequency band energy proportion and burst signal count rate; S3, constructing the features extracted in step (S2) into a defect-sensitive fusion vector and inputting the pre-trained adaptive weighted deep neural network model; S4, real-time identification of four types of defects of pores, cracks, incomplete fusion and interface brittle compound layer exceeding the standard based on the probability distribution output by the model. 2.The method according to claim 1, wherein, The sensor system configuration of step (S1) includes: the visible light high-speed camera collects at a rate of not less than 5000 frames / s, and uses an 808 nm filter to suppress plasma radiation interference; the spatial resolution of the infrared thermal imager is higher than 0.1 mm / pixel, and is clock-synchronized with the visible light camera through hardware triggering; the acoustic emission sensor is rigidly coupled to the waveguide rod within 10 mm from the weld; and the sampling time deviation of the three sensors is controlled within 1 μs. 3.The method of claim 1, wherein The molten pool contour segmentation in step (S2) specifically includes: a U-Net network architecture embedded with a spatial attention module is used, the encoder thereof extracts multi-scale features based on a ResNet34 backbone network, prior knowledge of the reflectivity difference of copper and aluminum materials is fused in the decoding stage, the segmentation result is corrected by threshold constraint, and finally the time-series sequence of molten pool width oscillation amplitude ΔW and rear solidification angle θ is output.

4. The method of claim 1, wherein the method is characterized by: The interface temperature gradient calculation in step (S2) includes: accurately marking the copper-aluminum heterogeneous material interface coordinate set in the infrared thermal image, expanding a 0.5 millimeter area on both sides of the interface normal as a temperature gradient calculation area, and calculating the temperature change amount per unit distance at a frequency of 100 Hz as a standardized gradient value 5. The method of claim 1, wherein the method is characterized by: The adaptive weighted deep neural network model of step (S3) includes a feature weighting layer, a multi-branch feature fusion module and a classification output layer, the feature weighting layer dynamically allocates weight coefficients according to the contribution of features to defect identification, the weight adjustment is based on the joint optimization of loss function gradient and feature information entropy, the multi-branch feature fusion module uses a one-dimensional convolutional neural network to process acoustic emission features, a long short-term memory network to analyze temperature time-series features, and a three-dimensional convolutional neural network to extract molten pool space-time features, and the classification output layer generates probability values of four types of defects through a Softmax function.

6. The method according to claim 5, wherein The model training adopts a cross-material transfer learning strategy, pre-trains the network infrastructure using a stainless steel welding dataset, fine-tunes the fully connected layer parameters on a copper-aluminum heterogeneous material specialized dataset, and fuses Focal Loss and distribution alignment metrics in the loss function to balance class imbalance and enhance domain adaptability.

7. The method of claim 1, wherein the method is characterized by: The real-time closed-loop control is also included, when the probability of porosity defects exceeds 80%, the pulse modulation control strategy of laser power is triggered, when the probability of interface brittle compound layer exceeds 75%, the mixing ratio of helium and argon is automatically adjusted, and the total delay time from defect identification to execution of control instruction is less than 50 milliseconds. 8.The method of claim 1, wherein, The acoustic emission feature extraction in step (S2) includes: using Daubechies wavelet basis for 5-layer wavelet packet decomposition, and calculating the energy ratio K of two characteristic frequency bands of 62.5-125 kHz and 125-250 kHz, when the energy ratio K is greater than 2.5 and the burst signal count rate exceeds 200 times / s, the high-risk crack warning is activated. 9.The method of claim 1, wherein, The copper-aluminum interface reaction knowledge base is constructed: the thermodynamic generation parameters of copper-aluminum alloy intermetallic compounds are pre-stored, and the interface reaction degree value is inversed based on real-time temperature field data, when the inversed value exceeds the preset threshold, the identification priority of interface brittle compound layer defects is improved.

10. A laser welding quality monitoring system for implementing the method of claims 1-9, characterized by It includes: Multimodal sensing unit: visible light camera, infrared thermal imager and acoustic emission sensor integrated with explosion-proof housing; Real-time processing unit: embedded industrial computer with graphics processor, used for executing feature extraction and defect identification algorithms; Human-computer interaction unit: visual interface displaying molten pool dynamics, temperature field distribution and defect probability; Control execution unit: feedback controller connected with laser generator and gas supply device.

Images