Pole piece welding defect detection device and method based on air coupled ultrasonic deep learning

Abstract

The present application relates to the technical field of nondestructive testing and artificial intelligence, and particularly relates to a kind of pole piece welding defect detection device and method based on air coupled ultrasonic deep learning, device includes air coupled ultrasonic detection module, multi-parameter signal processing and imaging module, deep learning defect identification module and online sorting control module, wherein, deep learning defect identification module contains MSDBNet model trained in advance.The welding defect detection device and method of the present application are suitable for the online full inspection of ultrasonic wave metal welding joint between tab and current collector (aluminum foil / copper foil) in the manufacturing process of lithium ion power battery cell, and can simultaneously realize the positioning, qualitative and level evaluation of seven typical defects such as internal porosity, crack, incomplete penetration, false welding, weak connection, slag inclusion and welding through.

Description

Technical Field

[0001] This invention relates to the fields of nondestructive testing and artificial intelligence, specifically to a device and method for detecting electrode welding defects based on air-coupled ultrasonic deep learning. Background Technology

[0002] As the core energy storage component of new energy vehicles, the safety, energy density, and cycle life of power batteries directly determine the overall performance of the vehicle. In the manufacturing process of power battery cells, ultrasonic metal welding between the tabs and current collectors is a critical step. The weld joint serves the dual function of conducting internal current and providing mechanical connection. Defects such as porosity, cracks, incomplete penetration, or weak welds in the weld joint can lead to a sharp increase in contact resistance, excessive local temperature rise, and potentially cause thermal runaway or even fire and explosion, resulting in serious safety accidents. Therefore, real-time, comprehensive, and reliable non-destructive testing of the weld joints of each cell under production line conditions is a rigid requirement for quality control in power battery manufacturing.

[0003] However, existing welding defect detection technologies in the power battery electrode inspection scenario consistently face a fundamental contradiction: the inability to simultaneously achieve "internal defect detectability, non-contact drying conditions, and online real-time full inspection speed." The specific existing technologies and their shortcomings are as follows:

[0004] If a manual visual inspection solution is adopted, the human eye has limited resolution and cannot reliably identify sub-millimeter defects. Furthermore, the attention of the inspector decreases after a long period of continuous inspection, which will lead to a sharp increase in the missed inspection rate. Manual visual inspection cannot meet the production capacity requirements and can only be carried out by sampling, which poses a risk of defective products being released. Different inspectors, or even the same inspector at different times, have inconsistent judgment standards, and the test results are not reproducible and lack data traceability.

[0005] If a machine vision inspection solution is adopted, and surface defects are identified through image processing algorithms or artificial intelligence models, there are problems such as severe reflection interference, inability to detect internal and subsurface defects, difficulty in covering the diverse morphologies of defects by traditional algorithms, and difficulty in identifying minute defects.

[0006] While contact ultrasonic testing can detect internal defects, it presents three insurmountable engineering problems: First, it requires a coupling agent, which severely conflicts with the dry and clean environment requirements of battery production lines, and residual coupling agent will contaminate the electrode sheets and affect subsequent processes; second, contact testing requires point-by-point scanning, resulting in slow testing speed that cannot match the second-level production line cycle time; and third, inconsistent contact pressure leads to fluctuations in acoustic coupling state, resulting in poor repeatability of test results.

[0007] If a conventional air-coupled ultrasonic testing scheme is adopted, using air as the coupling medium, there are two technical bottlenecks: First, the acoustic impedance mismatch between the piezoelectric crystal and the air is extremely severe, resulting in a sound wave reflectivity of over 99.99% at the crystal-air interface, extremely low transmission energy, and poor signal-to-noise ratio; Second, the air-coupled ultrasonic signal has high attenuation and high noise, making it difficult for traditional single-parameter detection methods to reliably extract defect information from the noise, and existing signal processing methods have a single dimension for defect characterization and insufficient discrimination.

[0008] Each of the aforementioned existing technical solutions can only satisfy one or two of the three conditions: "internal defect detectability, non-contact drying conditions, and online real-time full inspection speed." The remaining conditions suffer from fundamental, insurmountable defects. Therefore, there is an urgent need for a new detection technology capable of real-time full inspection of internal defects in welded joints under non-contact drying conditions at production line cycle speed. Summary of the Invention

[0009] This invention is made to solve the above-mentioned problems, and aims to provide a device and method for detecting electrode welding defects based on air-coupled ultrasonic deep learning. It is used for online full inspection of ultrasonic metal welding joints between electrode tabs and current collectors (aluminum foil / copper foil) during the manufacturing process of lithium-ion power battery cells. It can simultaneously realize the location, characterization and classification assessment of seven typical defects in the welding joint, including pores, cracks, incomplete penetration, false welds, weak connections, slag inclusions and weld penetration.

[0010] This invention provides a electrode welding defect detection device based on deep learning for air-coupled ultrasonic testing, used to detect defects in the welding area of ​​the electrode under test. It features the following components: an air-coupled ultrasonic detection module, used to emit an coded excitation signal to drive air-coupled ultrasonic wave emission, exciting a leaking Lamb wave to propagate in the electrode under test, and processing the leaking Lamb wave passing through the welding area to obtain a digital echo signal; a multi-parameter signal processing and imaging module, used to sequentially perform pulse compression and coherent averaging processing on the digital echo signal to generate a four-channel C-scan image with amplitude, time-of-flight, frequency centroid, and envelope width plots; and a deep learning defect recognition module, containing a pre-trained MSDBNet model, which inputs the four-channel C-scan image to the MSD. The MSDBNet model outputs defect identification results; the online sorting control module determines whether the tested electrode is qualified based on the defect identification results and controls the sorting action of the tested electrode. The MSDBNet model includes: an image branch for extracting spatial features from the four-channel C-scan image and outputting feature maps at four scales; a signal branch for extracting temporal features based on the original A-scan time-domain waveform retained in the four-channel C-scan image; a cross-modal bidirectional attention fusion module for bidirectional attention interaction fusion of spatial and temporal features to obtain fused features; a feature pyramid network for synchronously extracting and fusing multi-resolution features from the feature maps at four scales; and a joint task output head for outputting defect identification results based on the fused features and multi-resolution features.

[0011] The electrode welding defect detection device based on air-coupled ultrasonic deep learning provided by this invention may also have the following features: the air-coupled ultrasonic detection module includes: a transmitting transducer for transmitting air-coupled ultrasonic waves to the welding area of ​​the electrode under test, thereby exciting leakage Lamb waves to propagate in the electrode under test; a transmitting circuit for emitting a linear frequency modulated coded excitation signal to drive the transmitting transducer; a receiving transducer for receiving the leakage Lamb wave signal modulated by the welding area; a receiving circuit for amplifying and analog-to-digital conversion of the received leakage Lamb wave signal to obtain a digital echo signal; and a dual-probe mechanical scanning mechanism for driving the transmitting transducer and the receiving transducer to perform a two-dimensional scan along the welding area.

[0012] The electrode welding defect detection device based on air-coupled ultrasonic deep learning provided by the present invention may also have the following features: wherein the transmitting transducer and the receiving transducer are both installed on the same side of the welding area of ​​the electrode to be tested in a same-side oblique incidence layout, and the center line connecting the transmitting transducer and the receiving transducer forms a preset critical incidence angle with the normal direction of the electrode to be tested.

[0013] The electrode welding defect detection device based on air-coupled ultrasonic deep learning provided by the present invention may also have the following features: wherein the transmitting transducer is provided with three gradient acoustic matching layers, which are the first matching layer, the second matching layer and the third matching layer in sequence along the direction from the piezoelectric crystal side to the air side of the electrode under test, and the acoustic impedance of the three gradient acoustic matching layers decreases exponentially.

[0014] The electrode welding defect detection device based on air-coupled ultrasonic deep learning provided by this invention may also have the following features: the multi-parameter signal processing and imaging module includes: a pulse compression unit, which performs matched filtering on the echo signal to compress the coded excitation signal into a narrow pulse; a coherent averaging unit, which performs coherent superposition averaging on the echo signals from multiple repeated measurements; and a four-channel C-scan imaging unit, which generates four types of parameter maps based on the digital echo signal after pulse compression and coherent averaging, namely amplitude map, time-of-flight map, frequency centroid map, and envelope width map, and combines the above four types of parameter maps into a four-channel C-scan image, while retaining the original A-scan time-domain waveform corresponding to the four-channel C-scan image.

[0015] The electrode welding defect detection device based on air-coupled ultrasonic deep learning provided by the present invention may also have the following features: wherein the defect identification results include defect bounding box, segmentation mask, category probability, severity and risk score.

[0016] This invention also provides a method for detecting electrode welding defects based on air-coupled ultrasonic deep learning, characterized by the following steps: An air-coupled ultrasonic detection step, in which a transmitting circuit emits an coded excitation signal to drive a transmitting transducer to emit air-coupled ultrasonic waves towards the welding area of ​​the electrode under test, exciting leakage Lamb waves to propagate in the electrode under test; a receiving transducer receives the leakage Lamb wave signal passing through the welding area; and a receiving circuit amplifies and performs analog-to-digital conversion on the received leakage Lamb wave signal to obtain a digital echo signal; a multi-parameter signal processing and imaging step, in which a pulse compression unit sequentially performs matched filtering on the digital echo signal to compress the coded excitation signal into a narrow pulse; a coherent averaging unit performs coherent superposition and averaging of the echo signals from multiple repeated measurements; and a four-channel C-scan imaging unit generates four types of parameter maps based on the pulse-compressed and coherently averaged digital echo signals: amplitude map, time-of-flight map, frequency centroid map, and envelope width map; and combines these four types of parameter maps into a four-channel C-scan image, while retaining the original A-parameter signal corresponding to the four-channel C-scan image. The process involves several steps: First, scanning the time-domain waveform. Second, a deep learning defect identification step inputs a four-channel C-scan image to a pre-trained MSDBNet model. The image branch extracts spatial features from the four-channel C-scan image, outputting feature maps at four scales. The signal branch extracts temporal features from the original A-scan time-domain waveform corresponding to the four-channel C-scan image. A cross-modal bidirectional attention fusion module performs bidirectional attention interaction fusion of spatial and temporal features. A feature pyramid network simultaneously extracts and fuses multi-resolution features from the feature maps at four scales. Finally, a joint task output head outputs the defect identification result. Third, an online sorting control step determines whether the tested electrode is qualified based on the defect identification result and controls the sorting action on the tested electrode.

[0017] Compared with the prior art, the functions and effects of the present invention include:

[0018] The electrode welding defect detection device based on air-coupled ultrasound deep learning of the present invention can improve the signal-to-noise ratio by more than 20 dB because the air-coupled ultrasound detection module emits an coded excitation signal and the multi-parameter signal processing and imaging module performs pulse compression on the digital echo signal. By performing coherent averaging processing on the digital echo signal, the signal-to-noise ratio can be further improved, and the single-point detection time is shortened to the millisecond level. This overcomes the physical bottleneck of extremely low transmission energy caused by acoustic impedance mismatch between the piezoelectric crystal and the air, and realizes high signal-to-noise ratio air-coupled ultrasound excitation and reception.

[0019] In the welding defect detection device of the present invention, since the multi-parameter signal processing and imaging module generates amplitude map, time-of-flight map, frequency centroid map and envelope width map respectively through the four-channel C-scan imaging unit, and combines the above four types of parameter maps into a four-channel C-scan image, it constructs a multi-parameter joint characterization system, which can significantly improve the defect discrimination. In the MSDBNet model included in the deep learning defect recognition module, since the cross-modal bidirectional attention fusion module fuses the features of the four-channel C-scan image with the original A-scan time-domain waveform, the defect detection device and method of the present invention can comprehensively characterize the welding area from four physical dimensions: ultrasonic energy transmittance, interface delay, material nonlinearity and pulse scattering degree, so that each type of defect leaves a distinguishable feature fingerprint in four dimensions, which greatly improves the overall discrimination of input features for different defect types from the perspective of information theory. Furthermore, the visualization of the multi-channel C-scan image and the original A-scan time-domain waveform facilitates the review and traceability by inspection personnel.

[0020] Because the MSDBNet model employs a dual-branch heterogeneous architecture to process four-channel C-scan images and A-scan waveforms separately, and achieves adaptive dynamic fusion of the two features through a cross-modal bidirectional attention fusion module, and simultaneously extracts multi-resolution features at four scales through a feature pyramid network, the MSDBNet model possesses the ability to synchronously perceive defects across scales, from micropores below 0.3 mm to large-area incomplete weld penetration at the millimeter level. Furthermore, the MSDBNet model outputs defect identification results synchronously using a joint task output head, achieving a comprehensive three-in-one representation of defect location, characterization, and grading. Moreover, after knowledge distillation and compression, the MSDBNet model achieves an inference speed of less than 8 milliseconds per frame, with an overall detection cycle time of less than 2 seconds per cell, fully meeting the real-time full inspection requirements of power battery production lines at the second-level cycle time. The method of this invention integrates a deep learning model (MSDBNet model) with multi-modal and multi-scale information, ensuring both detection accuracy and real-time performance.

[0021] The electrode welding defect detection device based on air-coupled ultrasound deep learning of the present invention, because the transmitting transducer and the receiving transducer adopt the same-side dual-probe oblique incidence layout, can excite leakage Lamb waves by precisely controlling the incident angle. Taking advantage of the characteristics of Lamb waves that can propagate long distances in thin plates and are sensitive to internal defects, it can achieve high-sensitivity internal defect detection of internal and subsurface defects of welded joints under non-contact, dry conditions without coupling agent, completely avoiding the coupling agent contamination problem of contact ultrasound.

[0022] In the electrode welding defect detection device based on air-coupled ultrasonic deep learning of the present invention, because the transmitting transducer adopts a three-layer gradient acoustic matching layer, the acoustic impedance of the three matching layers decreases exponentially from about 3.0 MRayl on the piezoelectric wafer side to about 0.01 MRayl on the air side, realizing a smooth transition of acoustic impedance. This improves the acoustic wave transmittance by more than 15 dB compared with a single-layer matching structure, fundamentally solving the physical problem of extremely low transmission energy caused by acoustic impedance mismatch between the piezoelectric wafer and the air.

[0023] The electrode welding defect detection device and method based on air-coupled ultrasonic deep learning of the present invention outputs a defect identification result with a risk score of 0 to 100 as a continuous value, which can be directly connected to the quality classification management of the production line MES system; after knowledge distillation, the model can be deployed on the edge computing device of the production line without relying on cloud computing power. Therefore, the defect detection device and method of the present invention have good engineering practicality and interpretability. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the electrode welding defect detection device based on air-coupled ultrasonic deep learning in an embodiment of the present invention;

[0025] Figure 2 This is a schematic diagram of the structure of the three-layer gradient acoustic matching layer of the transmitting transducer in an embodiment of the present invention;

[0026] Figure 3 This is an acoustic impedance transition curve of the three-layer gradient acoustic matching layer of the transmitting transducer in an embodiment of the present invention;

[0027] Figure 4 This is a flowchart of the encoding excitation and pulse compression signal processing in an embodiment of the present invention;

[0028] Figure 5 This is a simulation result comparison of signal-to-noise ratio improvement at different processing stages of coding excitation, pulse compression, and coherent averaging in an embodiment of the present invention.

[0029] Figure 6 This is a time-domain comparison diagram of the A-scan echo signals before and after processing in an embodiment of the present invention;

[0030] Figure 7 This is a schematic diagram of the MSDBNet model architecture in an embodiment of the present invention;

[0031] Figure 8 This is a comparison chart of the detection performance of the MSDBNet model and the comparison model in the embodiments of the present invention;

[0032] Figure 9 This is a comparison of the training loss convergence curves of the MSDBNet model and the contrasting model in the embodiments of the present invention;

[0033] Figure 10 This is a simulation result diagram of the confusion matrix for the seven types of defects in the MSDBNet model in an embodiment of the present invention.

[0034] In the figure, 1. Air-coupled ultrasonic detection module; 11. Transmitting transducer; 12. Receiving transducer; 13. Transmitting circuit; 14. Receiving circuit; 15. Dual-probe mechanical scanning mechanism; 2. Multi-parameter signal processing and imaging module; 21. Pulse compression unit; 22. Coherent averaging unit; 23. Four-channel C-scan imaging unit; 3. Deep learning defect identification module; 30. MSDBNet model; 301. Image branch; 302. Signal branch; 303. Cross-modal bidirectional attention fusion module; 304. Feature pyramid network; 305. Joint task output head; 4. Online sorting control module; 100. Device. Detailed Implementation

[0035] To make the technical means, creative features, objectives and effects of the present invention easy to understand, the following embodiments, in conjunction with the accompanying drawings, provide a detailed description of the electrode welding defect detection device 100 based on air-coupled ultrasonic deep learning of the present invention.

[0036] Figure 1 This is a schematic diagram of the electrode welding defect detection device based on air-coupled ultrasonic deep learning in this embodiment.

[0037] like Figure 1 As shown, the present invention provides an electrode welding defect detection device 100 (hereinafter referred to as device 100) based on air-coupled ultrasonic deep learning, which is used to detect defects in the welding area of ​​the electrode under test. It has the following features: an air-coupled ultrasonic detection module 1, a multi-parameter signal processing and imaging module 2, a deep learning defect recognition module 3, and an online sorting control module 4.

[0038] The air-coupled ultrasonic testing module 1 includes a transmitting transducer 11, a transmitting circuit 13, a receiving transducer 12, a receiving circuit 14, and a dual-probe mechanical scanning mechanism 15.

[0039] Specifically, the transmitting circuit 13 is used to emit a linear frequency modulated coded excitation signal (center frequency 4 MHz, bandwidth 2 MHz, duration 50 microseconds) to drive the transmitting transducer 11; the transmitting transducer 11 is used to emit air-coupled ultrasonic waves to the welding area of ​​the electrode under test, thereby exciting the leakage Lamb wave to propagate in the electrode under test; the receiving transducer 12 is used to receive the leakage Lamb wave signal modulated by the welding area; the receiving circuit 14 is used to amplify (gain 60 dB) and perform analog-to-digital conversion (sampling rate 50 MHz, resolution 14 bits) on the received leakage Lamb wave signal to obtain a digital echo signal; the dual-probe mechanical scanning mechanism 15 is used to drive the transmitting transducer 11 and the receiving transducer 12 to perform a two-dimensional scan along the welding area.

[0040] Furthermore, both the transmitting transducer 11 and the receiving transducer 12 are installed on the same side of the welding area of ​​the electrode under test, using a same-side oblique incidence layout. The incident angle is adjusted to the critical incident angle θc≈15.1° for the leakage Lamb wave. The center line connecting the transmitting transducer 11 and the receiving transducer 12 forms a preset critical incident angle with the normal direction of the electrode under test.

[0041] Figure 2 This is a schematic diagram of the three-layer gradient acoustic matching layer of the transmitting transducer 11 in this embodiment. Figure 3 This is a diagram showing the acoustic impedance transition curve of the three-layer gradient acoustic matching layer of the transmitting transducer 11 in this embodiment.

[0042] like Figure 2 and Figure 3 As shown, the transmitting transducer 11 has three gradient acoustic matching layers, which are arranged sequentially from the piezoelectric crystal side to the air side along the electrode under test. The acoustic impedance of the three gradient acoustic matching layers decreases exponentially. The piezoelectric crystal of the electrode under test is made of PZT-5H piezoelectric ceramic with a thickness of 0.5 mm, a resonant frequency of 4 MHz, and an acoustic impedance of approximately 34 MRayl.

[0043] Specifically, the design principle of the three matching layers of the transmitting transducer 11 is as follows: For an n-layer matching layer structure, the theoretical optimal acoustic impedance Z of the i-th matching layer is... i It is determined by the following recursive formula:

[0044] Z i = (Z p × Z L ^i)^(1 / (i+1)) × (Z p / Z L )^(-i / (2(n+1))) .

[0045] Among them, Z p Z represents the acoustic impedance of the piezoelectric crystal.L Let be the acoustic impedance of the loading medium (air), n be the number of matching layers, and i be the matching layer number (i=1,2,...,n). When n=3, the theoretical optimal acoustic impedances of the three matching layers are as follows:

[0046] Z1 = (Z p × Z L )^(1 / 4) ≈ 3.0 MRayl

[0047] Z2 = (Z p × Z L ^2)^(1 / 4) ≈ 0.3 MRayl,

[0048] Z3 = (Z p × Z L ^3)^(1 / 4) ≈ 0.01 MRayl.

[0049] Z p =34 MRayl and Z L Substituting 0.0004 MRayl into the above formula, the theoretical optimal acoustic impedances of the three matching layers are approximately 3.0 MRayl, 0.3 MRayl, and 0.01 MRayl, respectively. Figure 3 As shown, the acoustic impedance of the three matching layers transitions smoothly along the thickness direction according to an exponential law, avoiding interface reflection caused by abrupt changes in acoustic impedance.

[0050] The thickness of each matching layer is designed to be one-quarter of a wavelength, that is:

[0051] d i = λ i / 4 = c i / (4 × f0) .

[0052] Where, d i Let λ be the thickness of the i-th matching layer. i Let c be the ultrasonic wavelength in the i-th matching layer. i Let fi be the sound velocity of the i-th matching layer material, and f0 be the center frequency of the transducer. The first matching layer 202 is made of epoxy resin-tungsten powder composite material (acoustic impedance approximately 3.0 MRayl, sound velocity approximately 2500 m / s) with a thickness of approximately 156 micrometers; the second matching layer is made of epoxy resin-hollow glass microsphere composite material (acoustic impedance approximately 0.3 MRayl, sound velocity approximately 1000 m / s) with a thickness of approximately 62 micrometers; the third matching layer is made of low-density polyurethane foam (acoustic impedance approximately 0.01 MRayl, sound velocity approximately 500 m / s) with a thickness of approximately 31 micrometers.

[0053] To further optimize the smoothness of the acoustic impedance transition, this embodiment introduces an exponentially varying acoustic impedance distribution between the three matching layers, the mathematical expression of which is:

[0054] Z(x) = Z1 × exp(-α × x) ,

[0055] α = ln(Z1 / Z) n ) / d total .

[0056] Where Z(x) is the acoustic impedance at position x along the thickness direction, Z1 is the acoustic impedance of the first matching layer, and Z... n For the acoustic impedance of the last matching layer, d total Let α be the total thickness of the matching layer and α be the exponential attenuation coefficient. This exponentially gradient distribution ensures a smooth and continuous transition of acoustic impedance within the matching layer, further reducing interface reflection and improving acoustic wave transmission efficiency. Simulation results show that the three-layer gradient matching layer structure improves acoustic wave transmission efficiency by approximately 15 dB or more compared to the single-layer matching structure, and by approximately 8 dB or more compared to the two-layer matching structure.

[0057] The multi-parameter signal processing and imaging module 2 includes a pulse compression unit 21, a coherent averaging unit 22, and a four-channel C-scan imaging unit 23. The pulse compression unit 21 performs matched filtering on the echo signal, compressing the coded excitation signal into a narrow pulse, thereby improving the signal-to-noise ratio by approximately 20 dB. The coherent averaging unit 22 is used to coherently superimpose and average the echo signals from multiple repeated measurements, further improving the signal-to-noise ratio by approximately 12 dB. The four-channel C-scan imaging unit 23 is used to generate four types of parameter maps based on the digital echo signal after pulse compression and coherent averaging: amplitude map, time-of-flight map, frequency centroid map, and envelope width map. These four types of parameter maps are then combined to form a four-channel C-scan image, while retaining the original A-scan time-domain waveform corresponding to the four-channel C-scan image.

[0058] Figure 4 This is a flowchart of the encoding excitation and pulse compression signal processing in this embodiment.

[0059] Specifically, to overcome the problem of low signal-to-noise ratio in air-coupled ultrasound, such as Figure 4 As shown, this embodiment employs chirp-coded excitation and pulse compression techniques. The mathematical expression for the chirp signal s(t) generated by the transmitting circuit 13 is:

[0060] s(t) = A(t) × cos(2π × (f0× t + K / 2 × t²)) ,

[0061] K = B / T.

[0062] Where A(t) is the signal envelope (usually using a Hanning window to reduce sidelobes), t is the instantaneous time (a continuous variable, defined in the domain 0 ≤ t ≤ T), f0 is the center frequency, B is the sweep bandwidth, T is the signal duration, and K is the frequency modulation slope. In this embodiment, f00 = 4 MHz, B = 2 MHz, T = 50 microseconds, and the corresponding time-bandwidth product BT = 100.

[0063] The pulse compression unit 21 performs matched filtering on the echo signal. The impulse response h(t) of the matched filter is the conjugate time reversal of the transmitted signal, and the formula is:

[0064] h(t) = s*(T - t).

[0065] The output y(t) after matched filtering is the convolution of the echo signal r(t) and the matched filter, as shown in the formula:

[0066] y(t) = r(t) * h(t).

[0067] The improvement in signal-to-noise ratio (SNR) after pulse compression processing is ΔSNR. pc Determined by the time-bandwidth product B and T, the formula is:

[0068] ΔSNR pc = 10 × log10(B × T) = 10 × log10(100) = 20 dB.

[0069] The coherent averaging unit 22 performs coherent superposition and averaging of the echo signals from N repeated measurements. Since the noise is a random signal and the measurements are uncorrelated, the signal-to-noise ratio improvement ΔSNR after coherent averaging is significant. avg for:

[0070] ΔSNR avg = 10 × log10(N).

[0071] In this embodiment, N=16, corresponding to ΔSNR avg =12 dB. Combining pulse compression and coherent averaging, the total improvement in signal-to-noise ratio in this embodiment is more than 32 dB, enabling the clear extraction of weak defect echoes that were originally submerged in noise.

[0072] Figure 5 This is a simulation result comparison of the signal-to-noise ratio improvement at different processing stages, including encoding excitation, pulse compression, and coherent averaging, in this embodiment. Figure 6 This is a time-domain comparison diagram of the A-scan echo signals before and after processing.

[0073] like Figure 5As shown, the original echo signal signal-to-noise ratio (SNR) was only about 5 dB. After pulse compression, the SNR increased to about 25 dB. After 16 coherent averaging operations, the SNR increased to about 37 dB. Figure 6 As shown, the defect echo is clearly identifiable.

[0074] The deep learning defect identification module 3 contains a pre-trained MSDBNet (Multi-Scale Dual-BranchFusion Network) model. It inputs a four-channel C-scan image and the initial A-scan time-domain waveform to the MSDBNet model 30 and outputs defect identification results, which include defect bounding boxes, segmentation masks, category probabilities, severity, and risk scores.

[0075] Figure 7 This is a schematic diagram of the architecture of MSDBNet model 30 in this embodiment.

[0076] like Figure 7 As shown, the MSDBNet model 30 includes an image branch 301, a signal branch 302, a cross-modal bidirectional attention fusion module 303, a feature pyramid network 304, and a joint task output head 305.

[0077] Specifically, image branch 301 takes a four-channel C-scan image as input and uses an improved ResNet-50 backbone network to extract spatial features. Image branch 301 includes four residual stages, outputting feature maps at scales P2, P3, P4, and P5, corresponding to downsampling factors of 4, 8, 16, and 32, respectively. Each residual stage consists of multiple stacked residual blocks, with 3×3 convolution, batch normalization, and ReLU activation function applied within each residual block.

[0078] Signal branch 302 is used to extract temporal features based on the original A-scan temporal waveforms preserved corresponding to the four-channel C-scan images, employing a hybrid architecture of a one-dimensional convolutional network and a Transformer. Signal branch 302 first extracts local features through three layers of one-dimensional convolutions with kernel sizes of 7, 5, and 3, and channel numbers of 32, 64, and 128, respectively. Then, it performs global temporal modeling through a built-in four-layer Transformer encoder. Each Transformer encoder layer includes a multi-head self-attention mechanism and a feedforward network, with 8 attention heads and a hidden layer dimension of 256. The temporal self-attention mechanism of signal branch 302 is calculated as follows:

[0079] Q = X × W Q K = X × W K V = X × W V ,

[0080] Attention(Q, K, V) = softmax(Q × K^T / ) × V.

[0081] Where X is the input feature sequence, W Q W K and W V d is a learnable projection matrix. k The dimension is the key vector. Q (Query) is the query vector, which is the input feature X projected onto the matrix W. Q The query features obtained after linear transformation are used to actively retrieve related information; K (Key) is the key vector, which is the input feature X after being projected through the matrix W. K The key features obtained after linear transformation are used as the index for retrieval; V (Value) is the value vector, which is the input feature X after being projected through the matrix W. V The value features obtained after linear transformation are used to carry the effective feature information to be extracted and fused.

[0082] The cross-modal bidirectional attention fusion module 303 is used to perform bidirectional attention interaction fusion of spatial features and temporal features to obtain fused features.

[0083] The Feature Pyramid Network 304 is used to simultaneously extract and fuse multi-resolution features from feature maps at four scales.

[0084] The joint task output head 305 is used to output defect identification results based on fused features and multi-resolution features, including a detection head, segmentation head, classification head, grading head, and scoring head.

[0085] The training of the MSDBNet model 30 adopts a two-stage strategy: the first stage uses WGAN-GP generative adversarial network to expand the training data, and the second stage uses knowledge distillation to compress the large model into a lightweight model.

[0086] Specifically, in the first stage, addressing the difficulty of obtaining a large number of defective samples in actual production lines, this embodiment employs a WGAN-GP (Wasserstein GAN with Gradient Penalty) generative adversarial network to generate synthetic defective samples to expand the training set. The loss function L of WGAN-GP... WGAN-GP for:

[0087] L WGAN-GP = E[D(x fake) ] - E[D(x real )] + λ × E[(||∇D(x̂)|| - 1)²],

[0088] Where D is the discriminator, x realFor real samples, x fake The synthesized sample is generated by the generator, x̂ is the interpolation point between the real sample and the synthesized sample, and λ is the gradient penalty coefficient (λ=10 in this embodiment). The synthesized defect samples generated by WGAN-GP are highly consistent with the real samples in terms of visual features and defect morphology, effectively alleviating the problem of insufficient training data.

[0089] In the second stage, knowledge distillation is used to compress the trained large model (teacher model) into a lightweight model (student model) for deployment on edge computing devices on the production line. The total loss function of knowledge distillation is L. total for:

[0090] L total = L hard + α × L KD + β × L feat ,

[0091] L KD = T² × KL(p T ^T || p S ^T),

[0092] Among them, L hard For the hard loss of the student model on the real labels, L KD For knowledge distillation loss, L feat The intermediate layer characteristic distillation loss is represented by α and β, which are weighting coefficients (α=0.7, β=0.3 in this embodiment), and T is the temperature coefficient (T=4 in this embodiment). T ^T and p S ^T represents the softening probability distribution of the teacher model and the student model at temperature T, respectively, and KL(·||·) is the KL divergence. After knowledge distillation, the number of parameters in the student model is compressed from 85M in the teacher model to 12M, the inference speed is improved from 35 milliseconds per frame to less than 8 milliseconds per frame, while the detection accuracy only decreases by about 1.2%.

[0093] The online sorting control module 4 determines whether the tested electrode is qualified based on the defect identification results and controls the sorting action of the tested electrode.

[0094] Specifically, the online sorting control module 4 includes a sorting mechanism. If the risk score R < 60, the tested electrode is deemed qualified; if R ≥ 60, it is deemed unqualified. The sorting mechanism performs sorting actions on the tested electrode according to the judgment result.

[0095] This embodiment also provides a method for detecting electrode welding defects based on air-coupled ultrasonic deep learning, including the following steps S1-S4.

[0096] S1, Air-coupled ultrasonic testing step: The transmitting circuit 13 sends an encoded excitation signal to drive the transmitting transducer 11 to emit air-coupled ultrasonic waves toward the welding area of ​​the electrode under test, which excites the leakage Lamb wave to propagate in the electrode under test. The receiving transducer 12 receives the leakage Lamb wave signal passing through the welding area. The receiving circuit 14 amplifies and performs analog-to-digital conversion on the received leakage Lamb wave signal to obtain a digital echo signal.

[0097] S2, Multi-parameter signal processing and imaging steps: The pulse compression unit 21 sequentially performs matched filtering on the digital echo signal to compress the coded excitation signal into a narrow pulse; the coherent averaging unit 22 performs coherent superposition averaging on the echo signals of multiple repeated measurements; and the four-channel C-scan imaging unit 23 generates four types of parameter maps based on the pulse-compressed and coherently averaged digital echo signals. The four types of parameter maps are amplitude map, time-of-flight map, frequency centroid map, and envelope width map. The above four types of parameter maps are combined into a four-channel C-scan image, while retaining the original A-scan time-domain waveform corresponding to the four-channel C-scan image.

[0098] S3, Deep Learning Defect Recognition Step: Input a four-channel C-scan image to a pre-trained MSDBNet model 30. Extract spatial features from the four-channel C-scan image through the image branch 301 and output feature maps at four scales. Extract temporal features from the original A-scan time-domain waveform corresponding to the four-channel C-scan image through the signal branch 302. Perform bidirectional attention fusion of spatial and temporal features through the cross-modal bidirectional attention fusion module 303. Simultaneously extract and fuse multi-resolution features from the feature maps at four scales through the feature pyramid network 304. Output the defect recognition result through the joint task output head 305.

[0099] S4, online sorting control step, determines whether the tested electrode is qualified based on the defect identification result, and controls the sorting action of the tested electrode.

[0100] Simulation Experiments and Result Analysis

[0101] To verify the effectiveness of the proposed solution, the applicant conducted comprehensive simulation experiments. The experimental dataset contains 12,000 welding samples, covering seven types of defects: porosity, cracks, incomplete penetration, weak welds, weak connections, slag inclusions, and burn-through. Of these, 8,000 were used for training, 2,000 for verification, and 2,000 for testing.

[0102] Figure 8 This is a comparison chart of the detection performance of the MSDBNet model 30 in this embodiment and the comparison model. Figure 9 This is a comparison chart of the training loss convergence curves of MSDBNet model 30 and the comparison model in this embodiment.

[0103] like Figure 8 As shown, the detection accuracy of the MSDBNet model in this embodiment is compared with four comparative models (YOLOv5, Faster R-CNN, ResNet+FPN, and single-branch ultrasound CNN) on the two metrics of mAP@0.5 and mAP@0.75. The results show that at mAP@0.5, the MSDBNet model achieves 93.7%, an improvement of 11.6 percentage points compared to the second-best model, Faster R-CNN (82.1%); at mAP@0.75, the MSDBNet model achieves 87.4%, an improvement of 19.1 percentage points compared to the second-best model, Faster R-CNN (68.3%). Among the other comparative models, ResNet+FPN has mAP@0.5 of 80.5% and mAP@0.75 of 65.2%, YOLOv5 has 78.3% and 62.5% respectively, and single-branch ultrasound CNN has 76.8% and 60.1% respectively, all significantly lower than the model proposed in this invention. Figure 9 As shown, the total loss of the MSDBNet model in this embodiment is consistently lower than that of the single-branch baseline model throughout 100 training iterations.

[0104] Simulation results of defect classification confusion matrix

[0105] Figure 10 This is a simulation result diagram of the confusion matrix for the seven types of defects in the MSDBNet model 30 in this embodiment.

[0106] like Figure 10 As shown, the MSDBNet model 30 achieved an overall accuracy of 95.8% for classifying seven defect categories on the test set. The recall rates for each defect category were as follows: porosity 94.2%, cracks 93.5%, incomplete penetration 96.8%, cold welds 95.1%, weak connections 96.3%, slag inclusions 94.7%, and burn-through 97.5%. The main confusion occurred between porosity and slag inclusions (approximately 3.5% of porosity was misclassified as slag inclusions) because they have similar morphologies in C-scan images; however, through four-channel multi-parameter joint characterization, the confusion rate has been significantly reduced.

[0107] Real-time test results

[0108] On the NVIDIA Jetson AGX Orin edge computing platform, the MSDBNet model 30, after knowledge distillation, has an inference speed of 7.8 milliseconds per frame. With the data acquisition and preprocessing time of about 1.2 seconds, the overall detection cycle time is less than 2 seconds per cell, which fully meets the real-time full inspection requirements of the power battery production line with a cycle time of seconds.

[0109] This embodiment, through systematic innovation in three aspects—transducer structure, excitation signal processing, and deep learning algorithm—has constructed a complete technical solution that meets engineering requirements under the three conditions of "internal defect detectability, non-contact drying conditions, and online real-time full inspection speed." It has achieved a comprehensive breakthrough from technical principles to engineering implementation in the field of non-destructive testing of welding defects in power battery electrode sheets, and has high technical novelty and significant industrial application value.

[0110] The role and effect of the embodiments

[0111] The electrode welding defect detection device 100 based on air-coupled ultrasonic deep learning in this embodiment can improve the signal-to-noise ratio by approximately 20% because the air-coupled ultrasonic detection module 1 emits an coded excitation signal and the multi-parameter signal processing and imaging module 2 performs pulse compression on the digital echo signal. The signal-to-noise ratio can be further improved by coherent averaging of the digital echo signal, reducing the single-point detection time to the millisecond level. Furthermore, because the multi-parameter signal processing and imaging module 2 generates amplitude, time-of-flight, frequency centroid, and envelope width maps respectively through the four-channel C-scan imaging unit, and combines these four types of parameter maps into a four-channel C-scan image, it constructs a multi-parameter joint characterization system, which can significantly improve defect discrimination. In the MSDBNet model 30 included in the deep learning defect recognition module 3, because the cross-modal bidirectional attention fusion module 303 fuses the features of the four-channel C-scan image with the original A-scan time-domain waveform, the defect detection device 100 and method of this invention can comprehensively characterize the welding area from four physical dimensions: ultrasonic energy transmittance, interface delay, material nonlinear characteristics, and pulse scattering degree. This leaves a distinctive feature fingerprint for each type of defect in all four dimensions, significantly improving the overall discrimination of input features for different defect types from an information theory perspective. The visualization of the multi-channel C-scan image and the original A-scan time-domain waveform facilitates verification and traceability by inspection personnel.

[0112] Because the MSDBNet model 30 adopts a dual-branch heterogeneous architecture to process the four-channel C-scan image and A-scan waveform respectively, and achieves adaptive dynamic fusion of the two features through the cross-modal bidirectional attention fusion module 303, and extracts multi-resolution features simultaneously at four scales through the feature pyramid network 304, the MSDBNet model 30 has the ability to synchronously perceive defects across scales from micropores below 0.3 mm to large-area incomplete weld penetration at the millimeter level. Moreover, the MSDBNet model 30 outputs the defect identification results simultaneously using a joint task output head 305, realizing a comprehensive three-in-one characterization of defect location, characterization, and classification. Furthermore, after knowledge distillation and compression, the inference speed of the MSDBNet model 30 reaches less than 8 milliseconds per frame, and the overall detection cycle is less than 2 seconds per cell, which fully meets the real-time full inspection requirements of the second-level cycle of the power battery production line, thereby achieving high-precision real-time detection.

[0113] The electrode welding defect detection device 100 based on air-coupled ultrasound deep learning in this embodiment has a dual-probe oblique incidence layout on the same side because the transmitting transducer 11 and the receiving transducer 12 are precisely controlled to excite the leakage Lamb wave. By utilizing the characteristics of the Lamb wave that can propagate over long distances in thin plates and is sensitive to internal defects, it can achieve high-sensitivity internal defect detection of internal and subsurface defects in welded joints under non-contact, dry conditions without coupling agent, completely avoiding the coupling agent contamination problem of contact ultrasound.

[0114] In the electrode welding defect detection device 100 based on air-coupled ultrasonic deep learning in this embodiment, because the transmitting transducer 11 adopts a three-layer gradient acoustic matching layer, the acoustic impedance of the three-layer matching layer decreases exponentially from about 3.0 MRayl on the piezoelectric wafer side to about 0.01 MRayl on the air side, realizing a smooth transition of acoustic impedance. This improves the acoustic wave transmittance by more than 15 dB compared to a single-layer matching structure, fundamentally solving the physical problem of extremely low transmission energy caused by acoustic impedance mismatch between the piezoelectric wafer and the air.

[0115] The electrode welding defect detection device 100 and method based on air-coupled ultrasonic deep learning in this embodiment outputs a defect identification result with a risk score of 0 to 100, which can be directly connected to the quality classification management of the production line MES system. After knowledge distillation, the model can be deployed on the edge computing device of the production line without relying on cloud computing power. Therefore, the defect detection device 100 and method of this invention have good engineering applicability and interpretability.

[0116] Those skilled in the art should understand that this invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to this invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A device for detecting welding defects in electrode sheets based on air-coupled ultrasonic deep learning, used to detect defects in the welding area of ​​the electrode sheet under test, characterized in that, include: An air-coupled ultrasonic testing module is used to send an encoded excitation signal to drive the air-coupled ultrasonic wave to be emitted, thereby exciting the leakage Lamb wave to propagate in the electrode under test, and to process the leakage Lamb wave passing through the welding area to obtain a digital echo signal. The multi-parameter signal processing and imaging module is used to sequentially perform pulse compression and coherent averaging on the digital echo signal to generate a four-channel C-scan image consisting of an amplitude map, a time-of-flight map, a frequency centroid map, and an envelope width map. The deep learning defect recognition module includes a pre-trained MSDBNet model. The four-channel C-scan image is input into the MSDBNet model, and the defect recognition result is output. The online sorting control module determines whether the tested electrode is qualified based on the defect identification result, and controls the sorting action of the tested electrode. The MSDBNet model includes: Image branch is used to extract spatial features from the four-channel C-scan image and output feature maps at four scales. The signal branch is used to extract timing features based on the original A-scan time-domain waveforms retained corresponding to the four-channel C-scan images; A cross-modal bidirectional attention fusion module is used to perform bidirectional attention interaction fusion on the spatial features and the temporal features to obtain fused features; A feature pyramid network is used to simultaneously extract and fuse multi-resolution features from the feature maps at four scales; A joint task output head is used to output the defect identification result based on the fused features and the multi-resolution features.

2. The electrode welding defect detection device based on air-coupled ultrasonic deep learning according to claim 1, Its features are: The air-coupled ultrasonic detection module includes: A transmitting transducer is used to emit air-coupled ultrasonic waves toward the welding area of ​​the electrode under test, thereby exciting the leakage Lamb wave to propagate in the electrode under test; A transmitting circuit is used to emit a linear frequency modulated coded excitation signal, thereby driving the transmitting transducer; A receiving transducer is used to receive the leaked Lamb wave signal modulated by the welding area; A receiving circuit is used to amplify and perform analog-to-digital conversion on the received leaked Lamb wave signal to obtain a digital echo signal; A dual-probe mechanical scanning mechanism is used to drive the transmitting transducer and the receiving transducer to perform a two-dimensional scan along the welding area.

3. The electrode welding defect detection device based on air-coupled ultrasonic deep learning according to claim 2, characterized in that: in, Both the transmitting transducer and the receiving transducer are installed on the same side of the welding area of ​​the electrode under test, using a same-side oblique incidence layout. The center line connecting the transmitting transducer and the receiving transducer forms a preset critical incidence angle with the normal direction of the electrode under test.

4. The electrode welding defect detection device based on air-coupled ultrasonic deep learning according to claim 2, characterized in that: in, The transmitting transducer has three gradient acoustic matching layers, which are arranged sequentially from the piezoelectric wafer side to the air side along the electrode under test. The acoustic impedance of the three gradient acoustic matching layers decreases exponentially.

5. The electrode welding defect detection device based on air-coupled ultrasonic deep learning according to claim 1, Its features are: The multi-parameter signal processing and imaging module includes: The pulse compression unit performs matched filtering on the echo signal to compress the encoded excitation signal into a narrow pulse. The coherent averaging unit is used to coherently superimpose and average the echo signals from multiple repeated measurements. The four-channel C-scan imaging unit is used to generate four parameter maps based on the digital echo signal after pulse compression and coherent averaging, namely amplitude map, time-of-flight map, frequency centroid map and envelope width map, and to combine the above four parameter maps into a four-channel C-scan image, while retaining the original A-scan time-domain waveform corresponding to the four-channel C-scan image.

6. The electrode welding defect detection device based on air-coupled ultrasonic deep learning according to claim 1, characterized in that: in, The defect identification results include defect bounding boxes, segmentation masks, category probabilities, severity, and risk scores.

7. A method for detecting electrode welding defects based on air-coupled ultrasonic deep learning, comprising using the electrode welding defect detection device based on air-coupled ultrasonic deep learning as described in any one of claims 1-6, characterized in that, Includes the following steps: The air-coupled ultrasonic testing step involves the transmitting circuit sending an coded excitation signal to drive the transmitting transducer to emit air-coupled ultrasonic waves toward the welding area of ​​the electrode under test, thereby exciting leakage Lamb waves to propagate in the electrode under test. The receiving transducer receives the signal of the leakage Lamb wave passing through the welding area, and the receiving circuit amplifies and performs analog-to-digital conversion on the received leakage Lamb wave signal to obtain a digital echo signal. The multi-parameter signal processing and imaging steps involve a pulse compression unit sequentially performing matched filtering on the digital echo signal to compress the coded excitation signal into a narrow pulse; a coherent averaging unit performing coherent superposition averaging on the echo signals from multiple repeated measurements; and a four-channel C-scan imaging unit generating a four-channel C-scan image with amplitude, time-of-flight, frequency centroid, and envelope width maps based on the pulse-compressed and coherently averaged digital echo signal. The deep learning defect identification steps are as follows: inputting the four-channel C-scan image into the pre-trained MSDBNet model; extracting spatial features from the four-channel C-scan image through the image branch and outputting feature maps at four scales; extracting temporal features based on the original A-scan temporal waveforms preserved in the four-channel C-scan image through the signal branch; performing bidirectional attention interaction fusion of the spatial and temporal features through a cross-modal bidirectional attention fusion module; simultaneously extracting and fusing multi-resolution features from the feature maps at four scales through a feature pyramid network; and outputting the defect identification result through a joint task output head. The online sorting control step determines whether the tested electrode is qualified based on the defect identification result and controls the sorting action of the tested electrode.

Images