Lithium battery laser welding detection method

Abstract

The invention discloses a lithium battery laser welding detection method, and belongs to the technical field of nondestructive testing. The method comprises the following steps: acquiring a three-dimensional path point sequence of a to-be-detected welding seam; controlling the transient thermal excitation unit to apply transient thermal excitation along the path; a high-speed multispectral imaging unit is used for synchronously collecting the thermal radiation attenuation process of the weld joint under at least two different spectral bands, and a space-time spectrum data cube is formed; and finally, inputting the data cube into an artificial intelligence model for defect identification. The system correspondingly comprises a three-dimensional visual sensor, a transient thermal excitation unit, a high-speed multispectral imaging unit and a synchronous control and data processing unit. According to the method, different types of welding defects such as air holes and incomplete fusion can be effectively distinguished by analyzing the heat diffusion characteristic difference under different spectral bands, the interference of uneven surface emissivity is inhibited, and the online detection accuracy and reliability of the three-dimensional complex welding seam are improved.

Description

Technical Field

[0001] This invention relates to the field of nondestructive testing technology, specifically to a testing method for laser welding of lithium batteries. Background Technology

[0002] The safety and lifespan of lithium batteries largely depend on the quality of their encapsulation and welding. Laser welding, due to its advantages of concentrated energy, high speed, and minimal deformation, has become a key process in lithium battery encapsulation. However, during high-speed welding, various internal and external defects such as porosity, cracks, and incomplete fusion are inevitably generated. These defects directly affect the battery's sealing performance and long-term reliability. Therefore, rapid and accurate online quality inspection of the weld seams is crucial.

[0003] Currently, infrared thermal imaging, as a non-contact non-destructive testing technology, has found some application in welding quality monitoring. Active infrared thermal imaging identifies defects by applying external thermal excitation to the workpiece and observing the temperature drop process on its surface. Defective areas exhibit abnormal temperature responses due to differences in thermophysical properties compared to the base material. However, existing active infrared thermal imaging detection methods still face some technical challenges when applied to scenarios such as lithium battery welds. First, conventional single-band infrared thermal imaging is susceptible to interference from uneven emissivity on the weld surface. Due to differences in oxidation, spatter, or crystallization states, the infrared emissivity at various points on the weld surface is not constant. This local variation in emissivity can be misinterpreted by single-band thermal imagers as a real temperature difference, resulting in false defect signals and leading to false alarms or missed detections. Second, relying solely on temperature attenuation information from a single spectral band makes it difficult to effectively classify defects.

[0004] Furthermore, welds in locations such as the top cover of lithium batteries are typically three-dimensional curves. Traditional fixed inspection systems struggle to maintain a constant detection distance and angle to the weld surface throughout the inspection process. This results in inconsistent signals acquired along the path, directly impacting the accuracy and repeatability of the inspection results, making it difficult to meet the requirements for automated online inspection of welds with complex trajectories. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a detection method for laser welding of lithium batteries, which solves the technical limitations of existing detection technologies when dealing with three-dimensional curved surfaces, uneven surface emissivity, and distinguishing between surface and internal defects.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for detecting laser welding of lithium batteries, the method comprising the following steps: Obtain the three-dimensional path point sequence of the weld seam to be tested in the lithium battery; Control a transient thermal excitation unit to apply transient thermal excitation to the weld under test according to the three-dimensional path point sequence; After the transient thermal excitation is applied, a high-speed multispectral imaging unit is used to simultaneously acquire the thermal radiation attenuation process of the weld surface under test in at least two different spectral bands, thereby obtaining a spatiotemporal spectral data cube. The spatiotemporal spectrum data cube is input into a pre-trained artificial intelligence model to identify defects and obtain the defect detection results of the weld to be tested.

[0007] In one specific implementation, the process of obtaining the three-dimensional path point sequence includes: scanning the area containing the weld to be tested using a three-dimensional vision sensor (e.g., a line laser profile scanner) to obtain three-dimensional point cloud data covering the area; subsequently, processing the three-dimensional point cloud data using algorithms, such as edge detection and surface fitting, to identify and fit the center line of the weld, and discretizing the center line to obtain the three-dimensional path point sequence.

[0008] In one specific implementation, the transient thermal excitation can be generated by a pulsed laser that integrates a scanning galvanometer system. The scanning galvanometer system rapidly scans the laser spot on a cross-section perpendicular to the weld direction, forming a uniform thermal excitation line.

[0009] By acquiring thermal radiation information across multiple spectral bands, the limitations of single-band detection can be overcome. According to Planck's law and Wien's displacement law, the peak wavelength of thermal radiation from an object differs at different temperatures. Normal and defective areas (such as porosity and lack of fusion) on a weld surface exhibit different cooling behaviors after being subjected to transient thermal excitation due to differences in their thermophysical properties. This difference in cooling behavior is not only reflected in the temperature value but also in the distribution and attenuation rate of radiant energy across different spectral bands.

[0010] In one specific implementation, the high-speed multispectral imaging unit achieves synchronous acquisition through a common optical path design. For example, the unit may include a thermal imager operating in the mid-infrared band (e.g., 3-5 μm) and a thermal imager operating in the long-infrared band (e.g., 8-12 μm). A dichroic beam splitter separates the thermal radiation from the weld under test. This beam splitter reflects the mid-infrared radiation to the first thermal imager while transmitting the long-infrared radiation to the second thermal imager, thereby achieving pixel-level synchronous imaging of the same target point in both bands.

[0011] Therefore, the collected spatiotemporal spectral data cube constitutes a four-dimensional function. ,in Let these be the spatial coordinates of a point on the weld surface. For the time following the incentive, The spectral band is specified. In one specific embodiment, the transient thermal excitation unit and the high-speed multispectral imaging unit are integrated and installed at the end of a multi-axis motion platform (such as a six-axis industrial robot). This multi-axis motion platform guides the probes of the two units to move precisely along the weld centerline based on the acquired three-dimensional path point sequence. Furthermore, the platform can also adjust the probe's orientation in real time based on the local normal direction information at the path points to maintain the incident angle of thermal excitation and the observation angle of thermal radiation constant or within a preset range throughout the detection process, thereby eliminating measurement errors caused by angle changes. When performing defect identification, the artificial intelligence model can directly process the original spatiotemporal spectral data cube. Alternatively, in another specific embodiment, the data can be preprocessed before inputting into the model to extract key physical features. The diffusion of heat in a medium follows the heat conduction equation. For the temperature decay of the surface of a one-dimensional semi-infinite object after being subjected to a transient thermal pulse, its surface temperature... Over time The change can be approximated as: ; in This refers to the thermal diffusivity of the material. The presence of defects alters the local thermal diffusivity. By analyzing each pixel in the spatiotemporal spectral data cube In each band Temperature decay curve below By fitting the data, one or more characteristic parameters that characterize the rate of decay can be calculated, such as the thermal relaxation time constant. Because different types of defects have different absorption and scattering characteristics for thermal radiation of different wavelengths, their calculated values ​​in different spectral bands vary. The values ​​will also show differences. These thermal relaxation time constants calculated under different spectral bands... , Multidimensional physical features are combined into a feature tensor, which is then input into an artificial intelligence model. This provides effective input for the model to identify and classify defects. The artificial intelligence model can employ deep learning structures such as convolutional neural networks (CNNs).

[0012] A second aspect of the present invention provides a detection system for laser welding of lithium batteries, the system comprising: A 3D vision sensor is used to acquire the 3D path point sequence of the weld seam to be tested in a lithium battery. Transient thermal excitation unit; A high-speed multispectral imaging unit is used to simultaneously acquire the thermal radiation attenuation process of the weld surface under test in at least two different spectral bands after the application of transient thermal excitation, so as to obtain a spatiotemporal spectral data cube. Synchronous control and data processing unit. This unit is connected to the above-mentioned components and is used to: control the transient thermal excitation unit to apply transient thermal excitation to the weld under test according to the three-dimensional path point sequence; and receive the spatiotemporal spectrum data cube, input it into a pre-trained artificial intelligence model for defect identification, and finally output the defect detection result of the weld under test.

[0013] This invention provides a method for detecting laser welding of lithium batteries. It has the following beneficial effects: 1. This invention obtains a spatiotemporal spectral data cube containing spatial, temporal, and spectral dimensions by simultaneously acquiring the thermal radiation attenuation process in at least two different spectral bands. This enables subsequent artificial intelligence models to comprehensively analyze the differences in heat diffusion characteristics under different spectral bands, thereby effectively distinguishing different types of welding defects such as porosity and lack of fusion, and suppressing detection interference caused by uneven emissivity of the weld surface, thus improving the accuracy and reliability of defect identification.

[0014] 2. This invention first acquires the three-dimensional path point sequence of the weld seam using a three-dimensional vision sensor, and then uses a multi-axis motion platform to guide the transient thermal excitation unit and the high-speed multispectral imaging unit to move along the path. This technical solution combines three-dimensional contour acquisition with the thermal imaging detection process, enabling the detection system to adapt to weld seams with complex curved surface features, achieving high-speed, automated online detection, and expanding the application scope of the technology. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the overall structure of the automated weld inspection system of the present invention; Figure 2 This is a schematic diagram of the weld seam three-dimensional contour acquisition process of the present invention; Figure 3 This is a schematic diagram of the internal optical path of the transient thermal excitation unit of the present invention; Figure 4 This is a schematic diagram of the optical path of the high-speed multispectral imaging unit of the present invention; Figure 5 This is a schematic diagram of the detection method of the present invention; Figure 6 This is a schematic diagram of the spatiotemporal spectrum data cube structure of the present invention.

[0016] Among them, 100 is the conveying and positioning unit; 200 is the transient thermal excitation unit; 300 is the high-speed multispectral imaging unit; 400 is the synchronous control and data processing unit; 410 is the synchronous controller; 420 is the data processor; 50 is the line laser; 60 is the scanning galvanometer system; and 70 is the dichroic beam splitter. Detailed Implementation

[0017] The technical solutions in 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.

[0018] The first step of the method of the present invention is the loading and positioning of the workpiece under test (S100). This step is used to transfer the lithium battery workpiece to be tested to the testing station and obtain the precise spatial position and orientation information of the weld seam to be tested in the coordinate system of the testing system. This step S100 may specifically include a workpiece conveying step (S101) and a weld seam coordinate acquisition step (S102).

[0019] Step S101, workpiece conveying. This step can be achieved through an automated conveyor line, for example, by placing the workpiece to be tested on a synchronous belt, pallet, or automated guided vehicle (AGV) and transporting it to a preset inspection station. In another embodiment, this step can also be completed by an industrial robot arm, which grasps the workpiece with its end effector gripper and precisely places it at the inspection position. For automated workpiece conveying, those skilled in the art can use existing mature automated integration solutions, the specific implementation of which is well known in the art and will not be described in detail here.

[0020] Step S102, Weld Coordinate Acquisition. After the workpiece arrives at the inspection station, the system needs to acquire the precise spatial coordinates and orientation of the weld to be inspected to guide subsequent thermal excitation and imaging operations. The weld coordinate acquisition step can be achieved through machine vision measurement or mechanical positioning.

[0021] In one embodiment, two-dimensional machine vision is used for measurement. A two-dimensional industrial camera fixed above the inspection station photographs the workpiece, acquiring a digital image of the weld area. The synchronous control and data processing unit 400 executes a preset image processing algorithm on the digital image, such as an edge detection algorithm or a template matching algorithm based on grayscale or color, to identify the set of pixels representing the weld in the image. By performing skeleton extraction, centerline fitting, or edge fitting on the pixel set, the path equation of the weld in the camera's two-dimensional image coordinate system can be obtained. Finally, using pre-calibrated camera intrinsic and extrinsic parameters and the transformation matrix between the camera and the system's motion axes, the two-dimensional image coordinates are converted into absolute coordinates in the system's three-dimensional spatial coordinate system, thereby completing the positioning.

[0022] In another specific embodiment, three-dimensional machine vision is used for measurement to obtain more complete position and orientation information of the weld. This three-dimensional machine vision system can be a line laser profile measurement sensor. During measurement, the sensor projects a laser line onto the weld surface and captures the deformation of the laser line caused by surface height undulations using an internal camera. Using triangulation principles, the system can calculate the three-dimensional coordinates of each point on the laser line in real time. By controlling the relative movement of the sensor or workpiece, the entire weld area is scanned, thereby obtaining high-density three-dimensional point cloud data covering the entire weld surface. This point cloud data not only defines the centerline trajectory of the weld but also contains the normal vector information of the weld surface, thus completely describing the spatial position and orientation of the weld. The data processor 420 processes this three-dimensional point cloud data, for example, by segmenting the weld area using curvature or height features and performing three-dimensional curve fitting on the segmented point cloud, ultimately obtaining a set of trajectory points describing the precise direction of the weld centerline in three-dimensional space.

[0023] In an alternative embodiment, the weld coordinate acquisition step S102 can also be achieved through high-precision mechanical positioning. The workpiece is placed in a precision fixture specifically designed for it, which strictly constrains the workpiece to a fixed, known spatial position through mechanical structures such as pin holes and positioning surfaces. In this case, the position and orientation of the weld to be tested relative to the system coordinate system are constant. This position and orientation coordinates can be obtained from design drawings or through a one-time offline calibration and stored as fixed parameters in the synchronous control and data processing unit 400. In subsequent batch inspections, the system can directly call the pre-stored coordinate data. The weld spatial coordinates and orientation data obtained through any of the above methods will be transmitted to subsequent steps as the basis for precise aiming and focusing by the transient thermal excitation unit 200 and the high-speed multispectral imaging unit 300.

[0024] The next step in the method is the detection parameter configuration S200. This step, performed before the actual excitation and acquisition operations, aims to set a set of suitable operating parameters for the subsequent detection process based on the material, size, surface condition, and expected defect type of the workpiece under test. These parameters will be loaded into the synchronous control and data processing unit 400 to precisely control the actions of all subsequent hardware.

[0025] In one embodiment, parameter configuration is performed manually by an operator through a human-machine interface (HMI) provided by the synchronous control and data processing unit 400. This interface graphically presents all configurable parameter items and allows the operator to input values ​​or select from preset options.

[0026] In an automated implementation, parameter configuration can be linked with the upper-level Manufacturing Execution System (MES). For example, the system automatically reads the barcode or QR code on the workpiece to be tested to identify its model information using a barcode reader equipped with the conveying and positioning unit 100. Subsequently, based on this model information, the system automatically retrieves and loads a set of process recipes bound to that model from a local database or a remote MES server. The process recipe is a set of preset testing parameters.

[0027] The detection parameters to be configured can be mainly divided into three categories: thermal excitation parameters, imaging parameters, and synchronization parameters. The thermal excitation parameters define the way energy is applied by the transient thermal excitation unit 200, which mainly includes pulse energy. and pulse width Pulse energy The instantaneous temperature rise of the weld surface directly affects the pulse width. Its value must be set in a balance between ensuring a sufficient signal-to-noise ratio in the thermal signal and avoiding any thermal damage to the workpiece. This affects the initial penetration depth of heat into the material. These parameters are ultimately translated into control commands for the driving circuitry of the excitation source (such as a pulsed laser).

[0028] Imaging parameters define the method by which the high-speed multispectral imaging unit 300 acquires images, and these mainly include the acquisition frame rate. Total collection time Integration time and spectral channel selection. Acquisition frame rate. The settings need to ensure sufficient sampling of the rapid decay process of the thermal signal to support accurate fitting of the subsequent thermal relaxation curve. Total acquisition time. This requires ensuring that the majority of the thermal relaxation process is recorded so that key features (such as the thermal relaxation time constant) can be stably calculated. The integration time affects image brightness and signal-to-noise ratio and needs to be adjusted based on the radiant intensity of the analyte. The selection of spectral channels determines the specific band combination for subsequent cross-spectral analysis.

[0029] Synchronization parameters define the temporal relationship between thermal excitation and imaging, with the trigger delay time being the core parameter. This parameter defines the time interval between the start of the thermal excitation pulse and the start of the camera acquiring the first frame of image. By setting different delay times, it is possible to focus on observing different stages of the thermal relaxation process, or to acquire background images before the excitation begins. Once configured, this entire set of parameters is fixed as the basis for the execution of this inspection until the next workpiece of a different type enters or a new manual instruction is received.

[0030] See attached document Figure 1 Appendix Figure 3 and appendix Figure 4After completing the parameter configuration, the method proceeds to the active excitation and transient data acquisition step (S300). This step is the core link in acquiring dynamic physical signals that reflect the internal quality of the weld, and it is precisely coordinated and executed by the synchronization control and data processing unit 400 according to the parameters set in step S200. This step S300 can be further decomposed into the thermal excitation implementation step S301, the transient image sequence acquisition step S302, and the timing synchronization coordination step S303.

[0031] Step S301, thermal excitation is implemented. Based on the weld coordinates and attitude data obtained in step S100, the synchronous control and data processing unit 400 controls the excitation beam of the transient thermal excitation unit 200 to precisely align with the weld area to be tested. Subsequently, the unit 400, according to the thermal excitation parameters (pulse energy) set in step S200... and pulse width The transient thermal excitation unit 200 sends a trigger command to the transient thermal excitation unit 200. The excitation source (e.g., a pulsed laser) of the transient thermal excitation unit 200 emits a short energy pulse, which passes through its internal beam shaping system to form a transient thermal field that uniformly covers the target area on the weld surface, causing the surface and subsurface temperatures of the weld to rise rapidly without damage.

[0032] Step S302, transient image sequence acquisition. Simultaneously with or within a very short time after the thermal excitation, the synchronous control and data processing unit 400 instructs the high-speed multispectral imaging unit 300 to begin operation. The high-speed multispectral imaging unit 300 operates according to the imaging parameters (acquisition frame rate) set in step S200. Total collection time (Integration time and selected spectral channels) to perform continuous, high-speed image acquisition of the weld area undergoing thermal relaxation (i.e., natural cooling).

[0033] In one specific embodiment, the imaging unit uses a high-speed rotating filter wheel synchronized with the camera's acquisition frames to switch different spectral channels between consecutive image frames, thereby acquiring the thermal radiation attenuation process across multiple spectral bands in a single acquisition. In another embodiment, this multispectral acquisition function can also be achieved through an imaging unit with a built-in beam splitting system. This system utilizes dichroic mirrors to separate light of different bands and simultaneously project it onto multiple independent focal plane arrays, thus achieving instantaneous synchronous acquisition of multispectral images. The final output of this step is a four-dimensional spatiotemporal spectral data cube, which can be represented as... ,in Image pixel coordinates, For the spectral band, This is relative to the start time of acquisition. The data cube is physically represented as a series of image files arranged in chronological order or a block of memory data, where each frame of the image itself contains multiple spectral channels.

[0034] Step S303, timing synchronization coordination. See attached document. Figure 4 The precise synchronization of the thermal excitation and data acquisition actions, measured in microseconds to nanoseconds, is key to this invention, as illustrated in the synchronous timing logic. The synchronization controller 410 within the synchronization control and data processing unit 400 serves as the timing reference for the entire system. During a typical detection cycle, the synchronization controller 410... At any given time, a trigger pulse is output to the transient thermal excitation unit 200, causing it to start a pulse with a width of The energy pulse. Simultaneously, the synchronization controller 410 sends a data acquisition trigger signal to the high-speed multispectral imaging unit 300, which is relative to... Each time has a configurable trigger delay time. The high-speed multispectral imaging unit 300 in The first frame of the image is captured at a certain time, and is... Continuous sampling at time intervals Frames, total acquisition time Specifically, it will trigger a delay time. Setting it to a tiny negative value allows the camera to acquire one or more frames of images before the thermal excitation pulse arrives. These images can then serve as a reference for background noise subtraction in subsequent data processing. This strict timing synchronization ensures the accuracy of the acquired spatiotemporal spectral data cube. It can accurately reflect the thermophysical properties of the weld material itself, rather than artifacts introduced by timing jitter.

[0035] See attached document Figure 1 The synchronization control and data processing unit 400 in the detection system disclosed in this invention is the command and computing core of the entire system. This unit is responsible for issuing precise timing control commands, receiving and processing high-speed acquired raw data, executing feature extraction and defect identification algorithms, and finally outputting structured detection results. In a specific embodiment, this unit 400 can be physically implemented by an industrial computer equipped with a high-performance computing card (such as a GPU) and a dedicated timing control card (such as an FPGA-based I / O card). Its functional logic mainly includes two modules: a synchronization controller 410 and a data processor 420.

[0036] The function of the synchronization controller 410 is to provide a unified, high-precision timing reference for the system, so as to achieve strict synchronization between the active excitation and transient data acquisition in step S300 of the method of the present invention. In a specific embodiment, the synchronization controller 410 is implemented by a field-programmable gate array (FPGA), which is capable of generating digital pulse signals with nanosecond-level precision. The controller 410 operates according to the synchronization parameters set in step S200, such as the trigger delay time. It generates and outputs at least two trigger signals: one to trigger the energy pulse of the transient thermal excitation unit 200, and the other to trigger the image acquisition of the high-speed multispectral imaging unit 300. Furthermore, when the imaging unit 300 employs a filter wheel scheme, the synchronization controller 410 is also responsible for outputting a stepper motor drive signal synchronized with the camera frame to ensure that the filter switches to the specified spectral channel position at the correct time.

[0037] The data processor 420 performs the data analysis and processing flow in the method of the present invention, including steps S400 to S600. In a specific embodiment, the function of the data processor 420 is implemented by a parallel computing program running on a graphics processing unit (GPU) to meet the need for real-time processing of large-scale data. The data processor 420 receives the spatiotemporal spectral data cube acquired and transmitted by the high-speed multispectral imaging unit 300. And perform the following processes in sequence: First, the raw data is preprocessed, including background subtraction using image frames acquired before excitation, application of non-uniformity correction (NUC) algorithms, and interpolation repair of detector defects, such as linear or bilinear interpolation. Second, transient spectral fingerprint extraction and feature map construction are performed (S400). For the time series of each pixel (x, y) in the image at each spectral band λ, the data processor 420 fits it to an exponential decay model using numerical optimization algorithms such as nonlinear least squares. This model can be expressed as: ; in, This is the radiation intensity signal at that point. For time, The initial amplitude of radiation intensity caused by thermal excitation. The thermal relaxation time constant is The background is a thermal radiation substrate. By solving this model, key physical features, especially the thermal relaxation time constant, are extracted for each pixel. Subsequently, the data processor 420 utilizes at least two different frequency bands (e.g., and The relaxation time constant of ) is used to calculate the transspectral relaxation ratio. Finally, the thermal relaxation time constant spectrum, transspectral relaxation ratio spectrum, and the spectrum of one or more bands in [the spectrum] will be combined. The initial thermal images at each time step are stacked and fused into a high-dimensional feature tensor, which serves as the input for subsequent processing.

[0038] Then, the AI-based defect identification and judgment S500 is executed. A pre-trained deep learning network model for pixel-level segmentation is embedded within the data processor 420. As a specific embodiment, this model can be a convolutional neural network based on the U-Net architecture, where the number of input layer channels matches the number of channels in the aforementioned high-dimensional feature tensor. The processor takes the high-dimensional feature tensor generated in the previous step as input and feeds it into the network model for a forward inference calculation, thereby outputting a pixel-level defect classification map with the same size as the original image. Finally, the execution result is output and archived (5600). The data processor 420 performs connected component analysis on the defect classification map, identifies independent defect regions, and calculates the area, centroid coordinates, and other geometric parameters of each defect. All this information is integrated and formatted to form a structured inspection report. This report can be displayed to on-site operators through a human-machine interface or uploaded to a Manufacturing Execution System (MES) for quality traceability and statistical analysis via communication interfaces such as Industrial Ethernet.

[0039] See attached document Figure 5 , Figure 5 This is a schematic diagram of an artificial intelligence model network architecture according to an embodiment of the present invention. After completing the construction of the high-dimensional feature tensor, the method flow enters the artificial intelligence-based defect identification and judgment step S500. This step aims to perform in-depth analysis on the multimodal features generated in the previous stage, which integrate static spectral information and dynamic thermophysical properties, and output pixel-level defect classification results.

[0040] In one specific embodiment, the artificial intelligence model used in this step is a deep learning network suitable for pixel-level segmentation tasks, such as a convolutional neural network based on the U-Net architecture. In another embodiment, other advanced semantic segmentation network architectures can also be used, such as the DeepLab series (e.g., DeepLabv3+), which utilizes atrous convolution to expand the receptive field without increasing computational parameters, and is also suitable for this type of defect segmentation task. The number of input layer channels of this network is specially designed to match the number of channels of the high-dimensional feature tensor generated in step S400, so that it can directly process data containing multiple physical features such as thermal relaxation time constants, transspectral relaxation ratios, and initial thermal images.

[0041] Reference Figure 5Taking U-Net as an example, the network model generally presents an encoder-decoder structure. The encoder consists of multiple consecutive downsampling modules, each typically containing one or two 3x3 convolutional layers, a non-linear activation function (such as ReLU), and a 2x2 max-pooling layer. Its function is to perform convolution and pooling operations on the input high-dimensional feature tensor layer by layer to extract deeper and more abstract semantic features. The decoder consists of multiple corresponding upsampling modules, each typically containing an upsampling operation (such as transposed convolution or bilinear interpolation) and several convolutional layers. This operation gradually restores the resolution of the deep semantic feature maps to their original size. A key structural element is the skip connection between the encoder and decoder. This connection directly transmits and fuses the shallow, high-resolution feature maps from the corresponding layers in the encoder path into the decoder path. This structure allows the network to simultaneously utilize deep defect category information and shallow defect contour details during the final decision, thereby achieving accurate segmentation of defect locations and boundaries. The last layer of the network model is a softmax activation function, whose output is a probability map of the same size as the input image, where each pixel... Each pixel is assigned a probability distribution representing its belonging to a predefined category. By performing an arqmax operation on the probability distribution of each pixel, the category with the highest probability value is selected as the final label for that pixel, thus forming a pixel-level defect classification map. Predefined categories can include "normal", "cold solder joint", "microcracks", "porosity", etc.

[0042] To enable the aforementioned network model to perform defect identification tasks, its network parameters need to be determined through a training process. The model training process includes the following steps: First, a training dataset containing a large number of labeled samples is established. Specifically, weld samples with known defects of different types and those without defects are collected, and a corresponding high-dimensional feature tensor is generated for each sample through the aforementioned steps (S100 to S400) of the method of this invention. Then, using offline analysis methods such as metallographic sectioning and X-ray flaw detection, the defect regions in each sample image are precisely labeled manually at the pixel level, forming ground truth label masks that correspond one-to-one with the feature tensors.

[0043] During training, feature tensors from the training dataset are batch-input into the network model, and the difference between the model's output predicted mask and the ground truth label mask is calculated. This difference is quantified by a loss function. Considering the class imbalance problem caused by the fact that the number of pixels in normal regions is much greater than the number of pixels in defective regions in practical applications, this loss function can be a combination function, such as a weighted combination of Dice loss and Focal loss. Dice loss is chosen because it can directly optimize the overlap (IoU) between the predicted and ground truth regions in the segmentation task and is insensitive to segmentation boundaries; while Focal loss reduces the weight of easily classified background pixels, making the model pay more attention to hard-to-identify defective pixels during training, thus jointly improving the detection ability of small defects. Then, an optimizer (such as the Adam optimizer) is used to update all weight parameters in the network model through backpropagation based on the gradient calculated by the loss function. This training process is iterated repeatedly until the value of the loss function converges to a stable range, marking the completion of model training.

[0044] In the actual execution of the defect identification and judgment step S500, the data processor 420 inputs the feature length of a weld to be tested into the already trained network model. The model performs a forward inference calculation and directly outputs the pixel-level defect classification map of the weld to be tested. This enables the automatic identification and classification of defects.

[0045] See attached document Figure 6 , Figure 6 This is a schematic diagram of a detection result display interface according to an embodiment of the present invention. After completing the defect identification and judgment based on artificial intelligence, the method flow enters the final result output and archiving step S600. This step is responsible for converting the pixel-level defect classification map generated in step S500 into structured information that is meaningful to operators and the production management system. This step can be further subdivided into defect area analysis S601, comprehensive quality assessment S602, and detection result presentation and archiving S603.

[0046] Step S601, Defect Region Analysis. The input for this step is the pixel-level defect classification map from step 5500. First, the data processor 420 performs image post-processing on the classification map, for example, applying morphological opening operations (erosion followed by dilation) to eliminate isolated noise points with excessively small areas caused by model misclassification. Then, the data processor 420 performs a connected component analysis algorithm based on 8-neighborhood recursion on the post-processed classification map, aggregating spatially adjacent pixels labeled as belonging to the same defect category into independent defect regions. For each identified independent defect region, the system calculates and extracts a series of quantification parameters, including but not limited to: the category of the defect region (e.g., cold weld, porosity), the area measured in pixels or physical units (e.g., square millimeters), the position of the geometric center (centroid) in the workpiece coordinate system, the dimensions of the circumscribed rectangle, and feature parameters describing its shape, such as the roundness calculated from the region's area and perimeter.

[0047] Step S602, Comprehensive Quality Assessment. Based on the quantitative parameters of all defect areas extracted in step S601, the data processor 420 makes an overall pass / fail judgment on the quality of the entire weld under test. This judgment is based on a set of pre-configured quality acceptance criteria. These criteria can be loaded as part of the process recipe in step S200, or set by an authorized engineer through a human-machine interface. As a specific embodiment, the acceptance criteria can be a series of logical rules that can be combined using Boolean logic (AND, OR, NOT), such as: "Any defect of the 'crack' category is not allowed"; "The maximum area of ​​a single 'porosity' defect must not exceed..." "; and "the total area of ​​all 'porosity' defects shall not exceed"; The data processor 420 compares the measured defect parameters with these rules one by one. If any rule is not met, the workpiece is judged to be unqualified.

[0048] Step S603, Presentation and Archiving of Detection Results. This step outputs the final results of the analysis and evaluation. In one embodiment, the results are visualized through a human-computer interaction interface (HMI). See [reference needed]. Figure 6 The interface can simultaneously display on one screen the original thermal image 10 of the weld to be tested, a pseudo-color image 20 overlaid with defect area outlines and labels, a detailed parameter list 30 containing all detected defects, and an overall assessment result 40 clearly marked "pass" or "fail". For the overlaid pseudo-color image 20, different categories of defects can be marked with different colors (e.g., red for cracks, blue for porosity) to provide intuitive differentiation. In a preferred embodiment, the user interface is interactive; for example, clicking on a row in the defect list 30 will highlight the corresponding defect area in the pseudo-color image 20, facilitating operator review.

[0049] Simultaneously, all inspection information, including the workpiece's unique identifier, inspection time, original spatiotemporal spectral data cube or its key feature maps, detailed parameters of each defect, and the final evaluation result, will be integrated into a structured inspection report. In one specific embodiment, this report can be formatted as a JSON or XML file. This report will be stored in a local database for future retrieval and can be uploaded in real time to the upper-level Manufacturing Execution System (MES) or Quality Management System (QMS) via an industrial Ethernet interface using standard TCP / IP protocol or specific industrial communication protocols (such as OPC-UA). As a configurable archiving strategy, the system can be set to archive only the structured inspection report file for workpieces judged as "qualified" to save storage space; while for workpieces judged as "unqualified," both the inspection report file and the complete original spatiotemporal spectral data cube will be archived to facilitate subsequent offline, more detailed engineering analysis.

[0050] See attached document Figure 1 and appendix Figure 2 The initial step of the method flow of this invention is the conveying and positioning of the weld (step S100), which is executed by the conveying and positioning unit 100 under the coordination of the synchronous control and data processing unit 400. The core task of this step is to convey the workpiece to be tested to the inspection station and accurately acquire the coordinates and attitude data of the weld to be tested in three-dimensional space, providing accurate guidance for subsequent excitation and acquisition steps. This step can be further decomposed into workpiece conveying and coarse positioning S101 and accurate three-dimensional contour acquisition and attitude calculation of the weld S102.

[0051] Step S101, workpiece conveying and coarse positioning. In one specific embodiment, the conveying and positioning unit 100 includes a conveyor belt driven by a servo motor. The workpiece to be tested is placed on the conveyor belt and conveyed to the detection area. When the workpiece travels to a preset detection station, a photoelectric sensor located next to the station detects the arrival of the workpiece and sends a trigger signal to the synchronous control and data processing unit 400. The unit 400 then instructs the conveyor belt to stop moving, thereby completing the coarse positioning of the workpiece within the detection field of view. In another embodiment, to accommodate slight positional and angular deviations that may exist on the conveyor belt, a two-dimensional machine vision sensor can be used. This sensor captures a top view of the workpiece and uses image recognition algorithms (such as template matching or feature matching) to locate a predefined feature of the workpiece (such as a corner or through hole), thereby calculating the precise position (X, Y) and rotation angle (Y) of the workpiece on the conveying plane. This location information is then passed to subsequent steps to compensate for positioning errors.

[0052] Step S102: Precise 3D contour acquisition and attitude calculation of the weld seam. After the workpiece comes to rest, a 3D vision sensor integrated within the conveying and positioning unit 100 begins operation to acquire precise spatial information of the weld seam. In one embodiment, the 3D vision sensor can be a line laser contour sensor, consisting of a line laser emitter and a camera. In another embodiment, the 3D vision sensor can also be a 3D camera based on structured light technology, which acquires the 3D point cloud of the entire field of view at once by projecting a pre-coded stripe pattern onto the workpiece and analyzing its deformation.

[0053] See attached document Figure 2 Taking a line laser profile sensor as an example, the sensor is mounted on a scanning axis composed of a high-precision linear module. Before use, the system needs to perform a hand-eye calibration process to accurately determine the transformation matrix between the coordinate system of the 3D vision sensor and the coordinate system of the subsequent actuators (such as the excitation and imaging units), which is the basis for achieving precise guidance. During operation, the line laser profile sensor projects a laser line 50 onto the workpiece surface and captures the laser line deformation caused by the undulations of the weld surface from a specific angle through its built-in camera. For each frame of image acquired by the camera, the processor inside the sensor calculates the 3D coordinates of hundreds of points on the laser line in real time based on the principle of triangulation, forming a profile line. The synchronous control and data processing unit 400 controls the scanning axis, driving the line laser profile sensor to perform a uniform linear scan along a direction perpendicular to the weld direction. During this process, the sensor continuously acquires a series of parallel profile lines, which, when combined, constitute high-density 3D point cloud data covering the entire area to be measured. .

[0054] Data processor 420 receives the 3D point cloud data and executes a weld identification and attitude calculation algorithm. In a specific embodiment, the algorithm first identifies the feature points of the weld by analyzing the gradient or curvature changes in the height (Z-axis) of the point cloud data. Since welds typically exhibit convex or concave structures, the curvature values ​​at their locations will significantly differ from those of flat base material areas. Specifically, the system calculates the local curvature of each point in the point cloud and identifies points with absolute curvature values ​​exceeding a preset threshold as candidate weld points. The system aggregates all identified candidate weld points to form a point set representing the weld centerline trajectory. Subsequently, the system uses a curve fitting algorithm, such as cubic spline interpolation or least squares, to fit this point set, thereby obtaining a spatial 3D curve that accurately describes the weld centerline. Finally, the spatial coordinates of each point on this 3D curve are... The tangent vector at that point (representing the local orientation or direction of the weld at that point) is extracted, forming an ordered sequence of path points. This sequence of path points will be passed to subsequent steps to guide the motion axes of the transient thermal excitation unit 200 and the high-speed multispectral imaging unit 300, ensuring that the excitation spot and the camera field of view can always be accurately aligned and follow the weld throughout the entire detection process.

[0055] See attached document Figure 1 and appendix Figure 3 In the method of the present invention, the step of actively stimulating the weld (step S200) is executed by the transient thermal excitation unit 200. The function of this unit is to apply a short thermal pulse with precisely controllable energy and duration on the weld path determined in step S100 according to the instructions of the synchronization control and data processing unit 400, so as to generate a momentary, unsteady temperature rise on the surface of the weld and its surrounding area.

[0056] In one specific embodiment, the core of the transient thermal excitation unit 200 is a high-power pulsed laser, such as a fiber laser, an Nd:YAG laser, or a CO2 laser. The laser wavelength is selected based on its ability to be efficiently absorbed by the material of the workpiece under test.

[0057] See attached document Figure 3 When a pulsed laser is used as the excitation source, its output energy pulses are transmitted via optical fiber and coupled to a scanning galvanometer system 60. This system includes two mirrors driven by high-precision motors and a focusing objective. In a preferred embodiment, to accommodate variations in the weld seam's height (Z-axis), the optical system also includes a dynamic focusing module. This module can be an electrically driven lens whose curvature is controlled by a data processor 420, ensuring that the laser spot maintains optimal focus at any position on the three-dimensional weld seam surface, thereby guaranteeing a constant energy density. By precisely controlling the deflection angles of the two mirrors through a synchronous control and data processing unit 400, the focused laser spot can move rapidly and flexibly within the scanning field of view, thereby forming an excitation region of a preset shape, such as a dot, line, or rectangle.

[0058] To achieve tracking of any three-dimensional weld seam, the entire transient thermal excitation unit 200 is mounted on a multi-axis motion platform, such as a six-axis industrial robot or a gantry-type three-axis (XYZ) Cartesian coordinate robot. The motion trajectory of this motion platform is guided by the weld seam centerline path point sequence calculated in step S100, ensuring that the center of the excitation area remains aligned with the weld seam centerline throughout the entire detection process, and that the normal direction of the excitation area (i.e., the laser incident direction) maintains a preset angular relationship (e.g., perpendicular incident) with the local normal direction of the weld seam surface. This is achieved by adjusting the end effector posture of the six-axis industrial robot.

[0059] When performing the excitation step S200, the specific process may include S201: setting excitation parameters. The synchronous control and data processing unit 400 sets and applies the excitation energy density according to the preset process formula. Key parameters, such as thermal conductivity and specific heat capacity, are set in advance through experiments or simulations based on the material properties of the workpiece under test (e.g., thermal conductivity, specific heat capacity), surface condition (e.g., color, roughness), and the required temperature response signal-to-noise ratio. The energy density can be determined by the following formula: ; in, The peak power of the laser. The duration of the laser pulse. This refers to the area of ​​the laser spot on the workpiece surface. By adjusting these parameters, it is possible to ensure a rapid and significant temperature rise on the workpiece surface while avoiding damage or melting of the material.

[0060] S202: Execute synchronous excitation. When the motion platform drives the transient thermal excitation unit 200 to move along the weld path to the designated detection point, the synchronous control and data processing unit 400 sends a high-precision time trigger signal to it. Upon receiving the signal, the transient thermal excitation unit 200 immediately outputs an energy pulse that conforms to the parameters set in step S201. This process is repeated continuously or in steps along the entire weld path. When using step-by-step execution, the motion platform moves between each detection point and pauses briefly, completing excitation and acquisition during the pause. When using continuous execution, the motion platform moves along the weld path at a constant speed, while the synchronous control unit triggers excitation pulses at predetermined spatial intervals based on the real-time position feedback from the motion platform encoder.

[0061] See attached document Figure 1 and appendix Figure 4 In the method of the present invention, step S300, which involves acquiring the thermal response, is performed by a high-speed multispectral imaging unit 300. The function of this unit is to immediately and synchronously record the thermal radiation attenuation process of the weld and its adjacent surface at high temporal resolution in multiple different spectral bands after the transient thermal excitation unit 200 applies a thermal pulse. The high-speed multispectral imaging unit 300 consists of two or more high-speed infrared thermal imagers.

[0062] See attached document Figure 4To ensure that the two thermal imagers can observe exactly the same field of view, their lens optical axes are precisely aligned. Thermal radiation entering the system first passes through a dichroic beam splitter 70. This beam splitter is designed to transmit light in one wavelength band (e.g., LWIR) while reflecting light in another wavelength band (e.g., MWIR). The transmitted and reflected beams then enter their respective thermal imagers, achieving spatial common-path acquisition. In another embodiment, this multispectral imaging function can also be achieved by placing a high-speed rotating filter wheel with different spectral bandpass filters in front of a single broadband high-speed infrared thermal imager. It should be noted that the filter wheel scheme is time-division multiplexing acquisition, not true synchronous acquisition. However, when the rotational speed of the filter wheel is much higher than the rate of change of the measured thermal phenomenon, time interpolation can be used to approximate synchronization in subsequent processing.

[0063] To achieve tracking of arbitrary three-dimensional weld seams, the entire high-speed multispectral imaging unit 300 and the transient thermal excitation unit 200 are integrated at the end of the same multi-axis motion platform (such as a six-axis robot) and maintain a fixed relative pose. Typically, the imaging unit's field of view is configured to immediately follow the excitation spot, continuously recording the cooling process of the just-excited area during scanning. The motion platform guides the imaging unit's field of view center to strictly follow the weld seam centerline path point sequence calculated in step S100. To ensure consistently clear images despite changes in weld seam height, the imaging unit's optical system may include an autofocus module. A crucial prerequisite is that each thermal imager must undergo radiometric calibration before use. This process establishes a precise mapping between the detector's output digital signal value and the corresponding radiation temperature by photographing a standard blackbody radiation source at different temperatures. This mapping is stored as a lookup table or polynomial function for converting the raw image data into actual temperature data in subsequent steps.

[0064] The acquisition step (S300) is precisely synchronized with the excitation step (S200), which can be described as step S301: S301: Synchronous Sequence Acquisition. The synchronization control and data processing unit 400 is key to achieving synchronization. Unit 400 first sends a trigger signal to the transient thermal excitation unit 200 to apply a thermal pulse. At the moment the excitation pulse ends, or after a microsecond-level, preset, extremely short start-delay, unit 400 simultaneously sends an acquisition trigger signal to all thermal imagers in the high-speed multispectral imaging unit 300 via a dedicated hardware trigger line. Upon receiving the signal, each thermal imager continuously captures a sequence of N frames at a preset frame rate (e.g., 1,000 frames per second). This process records the entire process of the weld surface naturally cooling from its heated peak temperature. The final output of this step is a set of time-series images corresponding to each spectral band, which can be represented as... ,in It is an index of the spectral bands. These are pixel coordinates. This represents the time elapsed since the start of data acquisition. This set of spatiotemporal spectral data cubes, encompassing spatial, temporal, and spectral dimensions, forms the basis for all subsequent analyses.

[0065] See attached document Figure 1 The synchronization control and data processing unit 400 is the central hub of the entire detection system, responsible for coordinating the operation of all other units in the system and executing all computational tasks from raw data acquisition to final result output. Physically, this unit can be an industrial computer (IPC) integrating a high-performance computing module and a real-time I / O interface. Logically, its functions can be divided into a synchronization controller 410 and a data processor 420.

[0066] The function of the synchronization controller 410 is to achieve precise temporal and spatial coordination of the entire system. In a specific embodiment, the synchronization controller 410 can be implemented by an I / O card based on a field-programmable gate array (FPGA) or a programmable logic controller (PLC) supporting hard real-time. It communicates and controls the servo motors in the conveying and positioning unit 100, the scanning axes of the 3D vision sensor, and the multi-axis motion platform carrying the excitation and imaging units via a deterministic industrial bus protocol (e.g., EtherCAT). During the execution of the inspection process, the synchronization controller 410 sends motion commands to the multi-axis motion platform according to the weld path point sequence planned by the data processor 420. At the same time, it continuously receives real-time feedback from the position encoder of the motion platform. When the motion platform reaches the predetermined inspection point on the path, the synchronization controller 410 sends trigger signals to the transient thermal excitation unit 200 and the high-speed multispectral imaging unit 300 simultaneously with microsecond-level precision through a dedicated hardware trigger circuit.

[0067] The data processor 420 is designed to process massive amounts of data and execute core algorithms. In one specific embodiment, the data processor 420 is a high-performance computing module, whose hardware configuration typically includes a multi-core central processing unit (CPU) and one or more graphics processing units (GPUs) for parallel computing acceleration. This processor connects to the high-speed multispectral imaging unit 300 via a high-bandwidth data interface (e.g., CameraLink, CoaXPress, or 10 Gigabit Ethernet). To ensure no data loss occurs at extremely high frame rates, a dedicated framegrabber can be configured in the data processor 420. This framegrabber is responsible for directly receiving image data from the thermal imager and storing it in the computer's main memory (RAM). In a preferred embodiment, to handle massive amounts of instantaneous data, the raw spatiotemporal spectral data cube is streamed directly from main memory to a high-speed, non-volatile storage medium, such as a RAID 0 array consisting of multiple solid-state drives (SSDs), forming a temporary data buffer. This ensures that no raw data is lost even if subsequent processing speed temporarily lags behind the acquisition speed.

[0068] The processing flow executed by the data processor 420 is scheduled by a main control software running on it. This main control software can be designed as a state machine, with states including: idle, locating, detecting, processing, and completed, ensuring the orderly and reliable execution of the entire complex process. The process includes: First, it receives and processes the 3D point cloud data from the conveying and positioning unit 100, executes weld seam recognition and attitude calculation algorithms, generates the 3D path point sequence required by the motion platform, and sends the sequence to the synchronization controller 410. Second, after acquisition, the data processor 420 reads the original spatiotemporal spectrum data cube from the temporary data buffer. .

[0069] Next, the data processor 420 performs computational tasks such as data preprocessing, feature extraction (step S400), and artificial intelligence defect identification (step S500). Specifically, pixel-level operations that can be parallelized on a large scale, such as radiometric scaling and feature extraction, as well as forward inference of the artificial intelligence model, are assigned to the GPU for maximum efficiency. Serial tasks, such as flow control, file I / O, and final report generation, are handled by the CPU.

[0070] Finally, the data processor 420 performs post-processing and comprehensive evaluation of the results (step 5600), generating a visualized inspection report and structured data files, and uploading the results to the Manufacturing Execution System (MES) via its network interface. In addition, this unit provides a human-machine interface (HM) through which operators can load process recipes, start or stop inspection cycles, monitor the real-time status of the system, and view historical inspection results.

Claims

1. A method for detecting laser welding of lithium batteries, comprising the following steps: Step 1: Obtain the three-dimensional path point sequence of the weld seam to be tested in the lithium battery using a three-dimensional vision sensor; Step 2: Control the transient thermal excitation unit to apply transient thermal excitation to the weld to be tested according to the three-dimensional path point sequence; Step 3: After the transient thermal excitation is applied, the thermal radiation attenuation process of the weld surface under test is simultaneously acquired using a high-speed multispectral imaging unit under at least two different spectral bands to obtain a spatiotemporal spectral data cube; Step 4: Input the spatiotemporal spectrum data cube into the pre-trained artificial intelligence model to perform defect identification and obtain the defect detection results of the weld to be tested.

2. The detection method for lithium battery laser welding according to claim 1, characterized in that, Step one includes: The three-dimensional vision sensor is controlled to scan the weld seam to be tested and acquire three-dimensional point cloud data covering the weld seam to be tested; The three-dimensional point cloud data is processed to identify and fit the weld centerline, thereby obtaining the three-dimensional path point sequence.

3. The detection method for lithium battery laser welding according to claim 1, characterized in that, The transient thermal excitation unit is a pulsed laser that includes a scanning galvanometer system.

4. The detection method for lithium battery laser welding according to claim 1, characterized in that, The high-speed multispectral imaging unit includes: One thermal imager operating in the mid-wave infrared band and one thermal imager operating in the long-wave infrared band; And a dichroic beam splitter for separating thermal radiation and directing it into two thermal imagers respectively.

5. The detection method for lithium battery laser welding according to claim 1, characterized in that, The transient thermal excitation unit and the high-speed multispectral imaging unit are jointly installed at the end of a multi-axis motion platform. The multi-axis motion platform guides the transient thermal excitation unit and the high-speed multispectral imaging unit to follow the weld seam to be tested according to the three-dimensional path point sequence.

6. The detection method for lithium battery laser welding according to claim 5, characterized in that, The multi-axis motion platform also adjusts the attitude of the transient thermal excitation unit and the high-speed multispectral imaging unit according to the local normal direction of the weld to be tested, so as to maintain the preset incident and observation angles.

7. The detection method for lithium battery laser welding according to claim 1, characterized in that, Before step four, the following are also included: The spatiotemporal spectrum data cube is processed to extract multidimensional physical features that characterize heat diffusion properties. These multidimensional physical features are then combined into a feature tensor, which is then input into the artificial intelligence model.

8. The detection method for lithium battery laser welding according to claim 7, characterized in that, The multidimensional physical characteristics include at least the thermal relaxation time constants calculated under different spectral bands.

9. The detection method for lithium battery laser welding according to claim 1, characterized in that, The artificial intelligence model is a convolutional neural network.

10. A detection system for lithium battery laser welding, comprising a detection method for lithium battery laser welding according to any one of claims 1-9, characterized in that, include: A three-dimensional vision sensor is used to acquire the three-dimensional path point sequence of the weld seam to be tested in the lithium battery. Transient thermal excitation unit; A high-speed multispectral imaging unit is used to simultaneously acquire the thermal radiation attenuation process of the surface of the weld under test in at least two different spectral bands after the transient thermal excitation is applied, so as to obtain a spatiotemporal spectral data cube. A synchronous control and data processing unit is used to: control the transient thermal excitation unit to apply transient thermal excitation to the weld to be tested according to the three-dimensional path point sequence; The spatiotemporal spectrum data cube is input into a pre-trained artificial intelligence model for defect identification, thereby obtaining the defect detection result of the weld to be tested.

Images