Power battery laser welding detection method integrating photoelectric-OCT in-situ sensing

Abstract

The present invention discloses a power battery laser welding detection method integrating photoelectric-OCT in-situ sensing, which relates to the field of battery welding detection technology. The method comprises the following steps: S1, integrating a photoelectric-optical coherence tomography in-situ sensor in a laser welding system, and arranging the optical probe and the laser welding head coaxially; S2, obtaining the position and real-time posture information of the laser welding head, combining the surface structure image in the optical tomography image, and tracking the welding seam trajectory in real time; S3, based on the frequency domain feature dissociation model, dissociating the feature information in the optical tomography image; S4, selecting the feature information as the input variable, and constructing a back propagation neural network model; S5, generating a power battery welding detection report based on the defect detection and classification recognition results. The present invention collects the optical tomography image in the welding process in real time, dynamically adjusts the welding position, improves the precise control capability of the welding process, and improves the detection accuracy and efficiency of welding defects.

Description

Technical Field

[0001] The present invention relates to the technical field of battery welding detection, and in particular to a power battery laser welding detection method integrating photoelectric-OCT in-situ sensing. Background Art

[0002] Power batteries are the power source for tools, mostly referring to batteries that power electric vehicles, electric trains, electric bicycles, and golf carts. Power batteries are the core components of new energy vehicles and an important direction for future energy transformation. They are mainly different from the starting batteries used to start automobile engines. They are mostly valve-sealed lead-acid batteries, open-top tubular lead-acid batteries, and lithium iron phosphate batteries.

[0003] The internal structure of power batteries is complex, and every manufacturing step from battery cells, modules to battery packs is crucial. Especially in the process of connecting various components, laser welding technology plays an irreplaceable role.

[0004] Laser welding, with its unique advantages such as flexibility, efficiency and precision, has become the preferred welding method in power battery production. It uses a high-energy laser beam to focus on the welding point in a very short time, so that the material quickly melts and forms a strong welded connection. This welding method is not only highly efficient, but also has a very low error rate, which provides a strong guarantee for the large-scale production of power batteries.

[0005] Laser welding inspection refers to the use of laser technology to weld battery components during the power battery manufacturing process, and the evaluation and control of welding quality through a series of inspection methods. The manufacturing quality of power batteries directly affects the safety, reliability and performance of the product. If there are welding defects inside the battery (such as pores, cracks, incomplete welding, etc.), it may cause local overheating during the battery charging and discharging process, and then cause thermal runaway, short circuit or even spontaneous combustion. In addition, poor welding quality may increase the internal resistance of the battery, cause local heating, and even cause welding point breakage during long-term use, leading to battery failure.

[0006] As the pace of power battery production accelerates, traditional detection methods (such as X-rays and ultrasound) may not be able to meet the needs of large-scale production in terms of detection speed, resulting in reduced production efficiency. Some detection methods (such as ultrasonic testing) may have insufficient resolution and accuracy when dealing with complex welded structures, making it difficult to detect tiny defects or weld details. In addition, current detection technologies often focus on the detection of a specific parameter, such as geometric dimensions or defect detection, but lack the ability to comprehensively evaluate multiple parameters. To achieve a comprehensive evaluation of welding quality, it is still necessary to integrate multiple detection methods and improve comprehensive detection capabilities. Summary of the invention

[0007] Based on this, it is necessary to provide a power battery laser welding detection method that integrates photoelectric-OCT in-situ sensing to address the above technical problems.

[0008] The present invention provides a power battery laser welding detection method integrating photoelectric-OCT in-situ sensing, the method comprising the following steps:

[0009] S1. Integrate the photoelectric-optical coherence tomography in-situ sensor in the laser welding system, arrange the optical probe coaxially with the laser welding head, and collect the optical tomography image in real time during the welding process;

[0010] S2, obtaining the position and real-time posture information of the laser welding head, combining the surface structure image in the optical tomography image, real-time tracking of the welding seam trajectory, and dynamically adjusting the welding position of the laser welding head;

[0011] S3, dissociating feature information in the optical tomography image based on the frequency domain feature dissociation model;

[0012] S4. Select feature information as input variables, construct a back propagation neural network model, and use the back propagation neural network model to perform defect detection and classification identification on the finished welding products of the power battery;

[0013] S5. Generate a power battery welding inspection report based on the defect detection and classification identification results.

[0014] Furthermore, obtaining the position and real-time posture information of the laser welding head, combining the surface structure image in the optical tomography image, real-time tracking of the welding seam trajectory, and dynamically adjusting the welding position of the laser welding head include the following steps:

[0015] S21, establishing the welding space coordinates of the laser welding system, and obtaining the coordinate position and real-time posture of the laser welding head in the welding space coordinates;

[0016] S22, generating a surface structure image of the optical tomography image by synthesizing multiple frames of the optical tomography image, and performing preprocessing and edge detection on the surface structure image to obtain an edge image to be welded;

[0017] S23, extracting ideal weld feature points by fitting the edge line to be welded in the edge image to be welded, and forming an ideal weld trajectory;

[0018] S24, tracking the ideal weld trajectory, calculating the position deviation between the laser welding head and the ideal weld in real time, and correcting the welding position of the laser welding head through feedback.

[0019] Furthermore, the surface structure image of the optical tomography image is generated by synthesizing multiple frames of the optical tomography image, and the surface structure image is preprocessed and edge detected to obtain the edge image to be welded, which includes the following steps:

[0020] S221, acquiring a plurality of frames of optical tomographic images captured continuously, aligning the plurality of frames of optical tomographic images using an image registration algorithm, and fusing and superimposing the images to form a surface structure image;

[0021] S222, performing noise removal, grayscale equalization and image smoothing processing of the surface structure image in sequence;

[0022] S223, performing a morphological opening operation on the surface structure image to eliminate sporadic pixel points in the background, and then performing a morphological closing operation to smooth the contour of the edge to be welded;

[0023] S224, using an edge enhancement algorithm to enhance the edge of the surface structure image, then using an edge detection algorithm to extract the edge to be welded in the surface structure image, using morphological operations to refine and correct the edge to be welded, and obtaining an image of the edge to be welded.

[0024] Furthermore, by fitting the edge line to be welded in the edge image to be welded, the ideal weld feature points are extracted, and the ideal weld trajectory is formed, which includes the following steps:

[0025] S231, selecting the edge point of the edge image to be welded as the origin and establishing the image coordinates;

[0026] S232, establishing a B-spline curve by setting control points and node vectors, and adjusting the B-spline curve by using the least square method to obtain a continuous and smooth edge line to be welded;

[0027] S233, setting ideal weld feature points in the edge line to be welded, wherein the ideal weld feature points include a starting point, an end point, a turning point, a curvature change point and an equidistant point;

[0028] S234, by changing the curve parameter values, searching for coordinate points that meet the ideal weld feature points on the edge line to be welded, and obtaining the pixel coordinates of the corresponding coordinate points to form an ideal weld trajectory.

[0029] Furthermore, tracking the ideal weld trajectory, calculating the position deviation between the laser welding head and the ideal weld in real time, and correcting the welding position of the laser welding head through feedback include the following steps:

[0030] S241, converting the pixel coordinates of the ideal weld trajectory into three-dimensional coordinates in the welding space coordinates by a hand-eye calibration method, and establishing an association relationship between the laser welding head and the weld trajectory;

[0031] S242, using a cubic non-uniform rational B-spline method to calculate the corresponding position of the laser welding head on the ideal weld trajectory, and finding the weld point closest to the laser welding head on the ideal weld trajectory;

[0032] S243, obtaining the position deviation between the laser welding head and the ideal weld trajectory by calculating the position difference between the laser welding head and the closest weld point, and correcting the position deviation of the laser welding head by using a controller.

[0033] Further, based on the frequency domain feature dissociation model, dissociating feature information in the optical tomography image includes the following steps:

[0034] S31, obtaining an optical tomographic image of a laser welding system in a historical welding process as a historical welding image, establishing an image training set, and constructing a frequency domain feature dissociation model;

[0035] S32. Define a combined loss function to guide the frequency domain feature dissociation model to characterize the frequency domain features, and perform supervised learning on the frequency domain encoding and dissociation results;

[0036] S33. Use the frequency domain feature dissociation model to extract different frequency domain codes and reconstruction codes in the real-time acquired optical tomography image, and through feature selection and dimensionality reduction processing, eliminate invalid features and redundant features, and dissociate to obtain feature information that meets the needs of power battery welding detection.

[0037] Furthermore, constructing the frequency domain feature dissociation model includes the following steps:

[0038] S311, construct a frequency domain dissociation encoder to extract frequency domain features of different levels from historical welding images and dissociate them into different frequency domain codes;

[0039] S312, constructing a frequency domain dissociation decoder to decode different frequency domain codes and generate welding fault images of different frequency domain sub-bands corresponding to the historical welding images;

[0040] S313, constructing a reconstruction decoder, generating a reconstruction code by performing feature fusion on different frequency domain codes, and then learning from the reconstruction code and generating a reconstructed image corresponding to the historical welding image.

[0041] Furthermore, the expression of the combined loss function is:

[0042]

[0043] Where L represents the combined loss function; L dwt represents the discrete wavelet change loss function; L r represents the regularization term loss function; Y represents the frequency domain sub-band image set formed in the historical welding image; represents the set of welding tomographic images after frequency domain decomposition; x represents the input optical tomographic image; y f Represents the frequency domain subband image corresponding to each frequency domain feature; represents the reconstructed image generated by the reconstruction encoder; f represents the frequency domain features extracted by the frequency domain dissociation encoder; It indicates that the welding fault images corresponding to different frequency domain features are generated by the frequency domain dissociation decoder; λ indicates the regularization coefficient.

[0044] Furthermore, the feature information includes low-frequency features, medium-frequency features and high-frequency features; among them, the low-frequency features represent the overall morphology and large-scale structural information of the weld, the medium-frequency features represent the medium-scale structural information of the weld, and the high-frequency features represent the surface detail information and edge information of the weld.

[0045] Furthermore, feature information is selected as input variables to construct a back propagation neural network model. The back propagation neural network model is used to perform defect detection and classification identification on the finished welding product of the power battery, including the following steps:

[0046] S41, defining the network architecture and model parameters of the back propagation neural network model, and defining feature labels based on the type of feature information as model input variables;

[0047] S42, using the leave-one-out method to divide the feature information into a training data set and a test data set, and performing model training and model tuning on the back propagation neural network model;

[0048] S43, inputting the characteristic information of the power battery obtained by real-time analysis into the back propagation neural network model, detecting defects existing in the welding process of the power battery, and identifying the defect type.

[0049] The beneficial effects of the present invention are:

[0050] 1. By using the photoelectric-optical coherence tomography (OCT) in-situ sensing technology, the optical tomography images of the welding process are collected in real time, combined with the welding head posture information and surface structure images, the welding position is dynamically adjusted, and the frequency domain feature dissociation model and back propagation neural network are used for defect detection and classification identification; it not only improves the precise control capability of the welding process, but also greatly improves the detection accuracy and efficiency of welding defects, making up for the shortcomings of poor real-time performance, low detection accuracy and inability to correct welding deviations in real time in traditional detection methods, and realizing the intelligence and automation of the welding process.

[0051] 2. By acquiring the posture information of the laser welding head in real time and combining it with the surface structure image generated by the optical tomography image, the precise tracking and dynamic adjustment of the weld trajectory during the welding process are achieved, thereby effectively correcting the position of the welding head, improving welding accuracy and stability, and meeting the needs of high-quality welding of power batteries.

[0052] 3. By constructing a frequency domain feature dissociation model, the multi-scale feature information in the optical tomography image can be effectively dissociated. The model extracts the key structural and defect features of the welding process through the steps of frequency domain encoding, decoding and reconstruction. The feature information not only provides data support for the real-time control of the welding process, but also provides a solid foundation for the evaluation of welding quality and defect detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0054] Figure 1 It is a flow chart of a power battery laser welding detection method integrating photoelectric-OCT in-situ sensing according to an embodiment of the present invention. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0056] See also Figure 1 , a power battery laser welding detection method integrating photoelectric-OCT in-situ sensing is provided, the method comprising the following steps:

[0057] S1. Integrate the optoelectronic-optical coherence tomography in-situ sensor into the laser welding system, arrange the optical probe coaxially with the laser welding head, and collect the optical tomography image during the welding process in real time.

[0058] Specifically, optoelectronic-optical coherence tomography (OCT) in-situ sensing is a non-contact optical imaging and sensing method based on optical coherence tomography technology. OCT in-situ sensing can provide high-resolution three-dimensional imaging and is mainly used for non-destructive testing of the internal structure of materials. OCT uses the principle of interferometry to penetrate the material using a near-infrared beam and constructs a three-dimensional image by analyzing the reflection information of the beam at different depths inside the material. This imaging technology can detect tiny structural changes inside the material, such as pores, cracks, and interlayer defects.

[0059] In addition, the main objects of power battery laser welding inspection include the following aspects:

[0060] 1. Welding of battery cells and tabs: The battery cell (core unit of the battery) is connected to its positive or negative tabs by laser welding. The quality of this welding point directly affects the conductivity and mechanical strength of the battery, and is the focus of testing.

[0061] 2. Welding between the tab and the busbar: The welding between the tab and the busbar is used to connect multiple cells into a battery pack. The welding quality affects the electrical performance, thermal management and structural stability of the battery pack.

[0062] 3. Welding of the battery shell and the sealing cover: The welding of the battery shell and the sealing cover ensures the sealing and structural integrity of the battery, preventing electrolyte leakage and the influence of the external environment on the internal part of the battery.

[0063] 4. Welding of battery modules and battery packs: The welding quality of metal connectors, busbars and other components in battery modules and battery packs is directly related to the performance and safety of the entire battery system.

[0064] S2. Obtain the position and real-time posture information of the laser welding head, combine it with the surface structure image in the optical tomography image, track the welding seam trajectory in real time, and dynamically adjust the welding position of the laser welding head.

[0065] In the description of the present invention, obtaining the position and real-time posture information of the laser welding head, combining the surface structure image in the optical tomography image, real-time tracking of the welding seam trajectory, and dynamically adjusting the welding position of the laser welding head include the following steps:

[0066] S21. Establish the welding space coordinates of the laser welding system, and obtain the coordinate position and real-time posture of the laser welding head in the welding space coordinates.

[0067] Specifically, an inertial measurement unit, including an accelerometer, gyroscope, and magnetometer, is integrated on the laser welding head to measure the acceleration, angular velocity, azimuth, and other information of the welding head. The IMU can provide the three-dimensional posture data (position, velocity, acceleration, tilt angle, azimuth, etc.) of the welding head in real time. Then, optical position sensors, such as laser trackers and optical encoders, are used to measure the position and motion trajectory of the welding head in space. These sensors can provide high-precision position information to help accurately locate the position of the welding head.

[0068] S22, generating a surface structure image of the optical tomography image by synthesizing multiple frames of the optical tomography image, and performing preprocessing and edge detection on the surface structure image to obtain an edge image to be welded.

[0069] In the description of the present invention, the surface structure image of the optical tomography image is generated by synthesizing multiple frames of the optical tomography image, and the surface structure image is preprocessed and edge detected to obtain the edge image to be welded, which includes the following steps:

[0070] S221, acquiring a plurality of frames of optical tomographic images captured continuously, aligning the plurality of frames of optical tomographic images using an image registration algorithm, and forming a surface structure image by superimposing and synthesizing the images.

[0071] S222, performing noise removal, grayscale equalization and image smoothing processing of the surface structure image in sequence.

[0072] S223, performing a morphological opening operation on the surface structure image to eliminate scattered pixels in the background, and then performing a morphological closing operation to smooth the contour to be welded.

[0073] S224, using an edge enhancement algorithm to enhance the edge of the surface structure image, then using an edge detection algorithm to extract the edge to be welded in the surface structure image, using morphological operations to refine and correct the edge to be welded, and obtaining an image of the edge to be welded.

[0074] S23, extracting ideal weld feature points by fitting the edge line to be welded in the edge image to be welded.

[0075] In the description of the present invention, the following steps are included to extract the ideal weld feature points by fitting the edge line to be welded in the edge image to be welded:

[0076] S231, selecting the edge point of the edge image to be welded as the origin and establishing the image coordinates.

[0077] S232, establishing a B-spline curve by setting control points and node vectors, and adjusting the B-spline curve using the least square method to obtain a continuous and smooth edge line to be welded.

[0078] S233, setting ideal weld feature points in the edge line to be welded, wherein the ideal weld feature points include a starting point, an end point, a turning point, a curvature change point and an equidistant point.

[0079] S234, by changing the curve parameter values, searching for coordinate points that meet the ideal weld feature points on the edge line to be welded, and obtaining the pixel coordinates of the corresponding coordinate points to form an ideal weld trajectory.

[0080] S24, tracking the ideal weld trajectory, calculating the position deviation between the laser welding head and the ideal weld in real time, and correcting the welding position of the laser welding head through feedback.

[0081] It should be explained that there may be deviations between the laser welding head and the ideal weld trajectory. The main reasons include the following aspects:

[0082] 1. In actual production, the assembly of welding workpieces often cannot be completely accurately aligned, and the position and angle of the workpieces may deviate from the design requirements. This assembly error will cause the weld position to be inconsistent with the predetermined trajectory. The deviation of the workpiece position is usually manifested as a lateral (X-axis, Y-axis) or vertical (Z-axis) deviation of the weld position, causing the laser welding head to deviate from the weld.

[0083] 2. During the welding process, due to the uneven heat input, the workpiece may be thermally deformed or distorted, resulting in changes in the weld position. The expansion or contraction of the workpiece material will also affect the actual position of the weld. Thermal deformation causes the weld trajectory to bend, undulate or other nonlinear deformations, and the laser welding head may deviate from the weld centerline.

[0084] 3. When the motion control system of the welding head performs weld tracking, the welding head may not be able to fully follow the slight changes in the weld trajectory due to system delays, insufficient sensor accuracy or algorithm errors. Dynamic response errors usually manifest themselves as the welding head failing to adjust its position in time at the bend or change point of the weld trajectory, resulting in deviations.

[0085] 4. The workpiece surface may be rough, uneven, dirty or have other surface defects, which may affect the actual position of the weld and cause deviations when the welding head follows the surface features. Deviations caused by surface irregularities may be up and down fluctuations or lateral deviations of the welding head along the weld line.

[0086] 5. The positioning system of the welding head, including sensors, servo systems, control algorithms, etc., may have inherent errors or inaccurate calibration, resulting in the position of the welding head being inconsistent with the expected position. The positioning error of the welding head may be a continuous offset or a random error, affecting the accurate positioning of the welding head.

[0087] Therefore, correcting the position of the laser welding head is to ensure that the welding head is always in the best welding position, that is, the laser beam of the welding head can be accurately aligned with the center of the weld to ensure the welding quality.

[0088] In the description of the present invention, tracking the ideal weld trajectory, calculating the position deviation between the laser welding head and the ideal weld in real time, and correcting the welding position of the laser welding head through feedback include the following steps:

[0089] S241. The pixel coordinates of the ideal weld trajectory are converted into three-dimensional coordinates in the welding space coordinates by the hand-eye calibration method, and the association relationship between the laser welding head and the weld trajectory is established.

[0090] S242. Calculate the corresponding position of the laser welding head on the ideal weld trajectory using the cubic non-uniform rational B-spline method, and find the weld point closest to the laser welding head on the ideal weld trajectory.

[0091] S243, obtaining the position deviation between the laser welding head and the ideal welding trajectory by calculating the position difference between the laser welding head and the closest welding point, and correcting the position deviation of the laser welding head by using a controller.

[0092] Specifically, the deviation between the actual position of the welding head and the expected position (closest to the weld point) is calculated, and then the position of the welding head is corrected through feedback from the control system so that it moves along the predetermined weld trajectory to ensure the accuracy and quality of welding.

[0093] S3. Based on the frequency domain feature dissociation model, the feature information in the optical tomography image is dissociated.

[0094] Specifically, the frequency domain dissociation feature reveals information such as the depth and material properties of the weld by converting the OCT signal into a frequency domain signal. These frequency domain features often reflect the physical changes inside the weld better than direct image features. Therefore, frequency domain feature extraction is an important part of the feature extraction process. Feature extraction uses frequency domain dissociation features to refine the key information in the welding process and form a feature vector for input into the neural network.

[0095] In the description of the present invention, based on the frequency domain feature dissociation model, dissociating feature information in the optical tomographic image includes the following steps:

[0096] S31. Obtain optical tomographic images of the laser welding system during the historical welding process as historical welding images, establish an image training set, and construct a frequency domain feature dissociation model.

[0097] In the description of the present invention, constructing a frequency domain feature dissociation model includes the following steps:

[0098] S311. Construct a frequency domain dissociation encoder to extract frequency domain features of different levels from historical welding images and dissociate them into different frequency domain codes.

[0099] Specifically, the frequency domain dissociative encoder uses discrete wavelet transform (DWT) to perform discrete wavelet transform on the input historical welding image. This process decomposes the image into multiple sub-bands, usually including low-frequency sub-bands (Approximation) and high-frequency sub-bands (details: Horizontal, Vertical, Diagonal). The decomposed sub-band information is encoded into different frequency domain codes, and different codes represent different frequency components of the image, such as edge information, texture information, etc. DWT can perform multi-level decomposition, that is, the low-frequency sub-bands are decomposed again to further obtain higher-level low-frequency features, which allows the different scale features of the welding image to be captured and encoded.

[0100] S312, constructing a frequency domain dissociation decoder, decoding different frequency domain codes, and generating welding fault images of different frequency domain sub-bands corresponding to historical welding images.

[0101] Specifically, the above different frequency domain codes are decoded by a frequency domain dissociation decoder, and the decoding process restores these codes into corresponding frequency domain sub-band images, which respectively represent the low-frequency, medium-frequency and high-frequency features in the original optical tomography image.

[0102] The different frequency domain sub-band images generated by the decoder correspond to the features of different scales in the welding tomographic image. The low-frequency sub-band image may show the overall shape of the weld, the medium-frequency sub-band image reveals the medium-scale structure of the weld, and the high-frequency sub-band image shows the fine features of the weld surface and edge.

[0103] S313, constructing a reconstruction decoder, generating a reconstruction code by performing feature fusion on different frequency domain codes, and then learning from the reconstruction code and generating a reconstructed image corresponding to the historical welding image.

[0104] Specifically, a reconstruction decoder is used to fuse the decoded different frequency domain codes. The fusion process generates a comprehensive reconstruction code by aggregating low, medium and high frequency features. The reconstruction code contains multi-scale feature information in the welding fault image.

[0105] The reconstruction code is learned and reconstructed images corresponding to the original optical tomographic images are generated. These reconstructed images integrate multi-scale feature information and represent the comprehensive characteristics of the weld area.

[0106] S32. Define a combined loss function to guide the frequency domain feature dissociation model to characterize the frequency domain features, and perform supervised learning on the frequency domain encoding and dissociation results.

[0107] In the description of the present invention, the expression of the combined loss function is:

[0108]

[0109] Where L represents the combined loss function; L dwt represents the discrete wavelet change loss function; L r represents the regularization term loss function; Y represents the frequency domain sub-band image set formed in the historical welding image; represents the set of welding tomographic images after frequency domain decomposition; x represents the input optical tomographic image; y f Represents the frequency domain subband image corresponding to each frequency domain feature; represents the reconstructed image generated by the reconstruction encoder; f represents the frequency domain features extracted by the frequency domain dissociation encoder; It indicates that the welding fault images corresponding to different frequency domain features are generated by the frequency domain dissociation decoder; λ indicates the regularization coefficient.

[0110] S33. Use the frequency domain feature dissociation model to extract different frequency domain codes and reconstruction codes in the real-time acquired optical tomography image, and through feature selection and dimensionality reduction processing, eliminate invalid features and redundant features, and dissociate to obtain feature information that meets the needs of power battery welding detection.

[0111] In the description of the present invention, feature information includes low-frequency features, medium-frequency features and high-frequency features. Among them, low-frequency features represent the overall shape and large-scale structural information of the weld, medium-frequency features represent the medium-scale structural information of the weld, and high-frequency features represent the surface detail information and edge information of the weld.

[0112] Specifically, feature information includes the following categories:

[0113] 1. Low frequency characteristics

[0114] Overall shape and profile: Low-frequency features usually represent the overall shape of the weld and large-scale structural information, such as the width and overall profile of the weld. These features can reflect the macroscopic morphology of the weld and ensure the uniformity and continuity of the weld on a large scale.

[0115] 2. Intermediate frequency characteristics

[0116] Mesoscale structure: The medium-frequency features reflect the mesoscale structure of the weld, such as the depth variation of the weld, the local thickness of the weld zone, the local melting degree of the material, etc. These features can reveal medium-scale defects that may occur during the welding process, such as insufficient local thickness or excessive melting.

[0117] 3. High frequency features

[0118] Surface details and edge information: High-frequency features are usually associated with detailed features of the weld surface texture, micro-cracks, pores, etc. These features are very important because they can reveal tiny defects that may exist on the weld surface, which, if not detected, may affect the long-term stability and reliability of the weld.

[0119] S4. Select feature information as input variables, build a back-propagation neural network model, and use the back-propagation neural network model to perform defect detection and classification identification on the finished welding products of the power battery.

[0120] In the description of the present invention, feature information is selected as an input variable, a back propagation neural network model is constructed, and defect detection and classification identification of the finished welding product of the power battery using the back propagation neural network model includes the following steps:

[0121] S41. Define the network architecture and model parameters of the back propagation neural network model, and define feature labels based on the type of feature information as model input variables.

[0122] S42. Use the leave-one-out method to divide the feature information into a training data set and a test data set, and perform model training and model tuning on the back propagation neural network model.

[0123] S43, inputting the characteristic information of the power battery obtained by real-time analysis into the back propagation neural network model, detecting defects existing in the welding process of the power battery, and identifying the defect type.

[0124] Specifically, by inputting low-frequency features, medium-frequency features, and high-frequency features into the back propagation neural network, various common defects in the power battery welding process can be identified and classified. Each feature reflects different physical phenomena in the welding process at different frequencies, thereby helping the network to identify and classify defects. The following is an embodiment of the specific defect types corresponding to these features and their detection and classification methods:

[0125] 1. Low frequency features - detecting defects

[0126] 1) Weld offset: The low-frequency feature can reflect the overall trajectory of the weld. If the actual position of the weld is offset from the designed position, the low-frequency feature will show the offset of the weld.

[0127] 2) Abnormal weld width: Through low-frequency characteristics, abnormal changes in weld width can be detected. For example, a weld width that is too wide or too narrow may indicate unstable welding power or inconsistent welding speed.

[0128] 3) Weld discontinuity: Low-frequency features can also reveal weld discontinuities, such as weld breaks and jumps, which usually appear as sudden width changes or absences in the weld morphology.

[0129] Classification method: After the low-frequency features are input into the back propagation neural network model, the back propagation neural network model can identify whether there are abnormalities in the overall morphology of the weld by analyzing these macro features, and classify them into defect categories such as offset, width abnormality, discontinuity, etc.

[0130] 2. Intermediate frequency characteristics - detection of defects

[0131] 1) Porosity: The medium frequency characteristics can detect the pores caused by gas inclusions inside the weld. These pores usually appear as frequency characteristic changes in a small local area.

[0132] 2) Microcracks: These cracks are difficult to detect directly through macroscopic features, but the medium-frequency features can capture the strength changes or discontinuities in local areas of the weld, indicating the possible presence of microcracks.

[0133] 3) Weld Porosity: Porosity within the weld is also a key target for detection through medium frequency features. These features can identify voids or defective areas within the weld.

[0134] Classification method: After the intermediate frequency features are input into the back propagation neural network model, the back propagation neural network model can detect the integrity of the internal structure of the weld by analyzing these features, and classify and identify internal defects such as pores, microcracks and porosity.

[0135] 3. High frequency features - detecting defects

[0136] 1) Surface cracks: High-frequency features can capture cracks on the weld surface caused by stress concentration or uneven cooling of the material. These cracks usually manifest themselves in abnormal changes in surface texture.

[0137] 2) Abnormal surface roughness: If the weld surface roughness is uneven, the high-frequency features can detect these changes, for example, the weld surface is too granular or unevenly melted.

[0138] 3) Oxidation or contamination: High frequency features can also identify oxide layers or contaminant residues on the weld surface, which usually appear as subtle changes in the weld surface and abnormal smoothness.

[0139] Classification method: After the high-frequency features are input into the back-propagation neural network model, the back-propagation neural network model can identify the abnormalities of the weld surface details and classify these abnormal details into categories such as surface cracks, roughness abnormalities, oxidation or contamination.

[0140] S5. Generate a power battery welding inspection report based on the defect detection and classification identification results.

[0141] In summary, with the help of the above technical solution of the present invention, by using the photoelectric-optical coherence tomography (OCT) in-situ sensing technology, the optical tomography images in the welding process are collected in real time, combined with the welding head posture information and surface structure image, the welding position is dynamically adjusted, and the frequency domain feature dissociation model and the back propagation neural network are used for defect detection and classification identification; not only the precise control ability of the welding process is improved, but also the detection accuracy and efficiency of welding defects are greatly improved, making up for the shortcomings of poor real-time performance, low detection accuracy and inability to correct welding deviations in real time in traditional detection methods, and realizing the intelligence and automation of the welding process. By acquiring the posture information of the laser welding head in real time and combining the surface structure image generated by the optical tomography image, the accurate tracking and dynamic adjustment of the weld trajectory during the welding process are realized, thereby effectively correcting the position of the welding head, improving the welding accuracy and stability, and meeting the needs of high-quality welding of power batteries. By constructing a frequency domain feature dissociation model, the multi-scale feature information in the optical tomography image can be effectively dissociated. The model extracts the key structural and defect features in the welding process through the steps of frequency domain encoding, decoding and reconstruction. The feature information not only provides data support for the real-time control of the welding process, but also provides a solid foundation for welding quality evaluation and defect detection.

[0142] It should be understood that, although the steps in the flowchart of the accompanying drawings are displayed in sequence as indicated by the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least a part of the steps in the flowchart of the accompanying drawings may include multiple sub-steps or multiple stages, and these sub-steps or stages are not necessarily executed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be executed in turn or alternately with other steps or at least a part of the sub-steps or stages of other steps.

Claims

1. A power battery laser welding detection method integrating photoelectric-OCT in-situ sensing, characterized in that: The method comprises the following steps: S1. Integrate the photoelectric-optical coherence tomography in-situ sensor in the laser welding system, arrange the optical probe coaxially with the laser welding head, and collect the optical tomography image in real time during the welding process; S2, obtaining the position and real-time posture information of the laser welding head, combining the surface structure image in the optical tomography image, real-time tracking of the welding seam trajectory, and dynamically adjusting the welding position of the laser welding head; S3, dissociating feature information in the optical tomography image based on the frequency domain feature dissociation model; S4. Select feature information as input variables, construct a back propagation neural network model, and use the back propagation neural network model to perform defect detection and classification identification on the finished welding products of the power battery; S5. Generate a power battery welding inspection report based on defect detection and classification identification results; The method of dissociating the characteristic information in the optical tomographic image based on the frequency domain characteristic dissociation model comprises the following steps: S31, obtaining an optical tomographic image of a laser welding system in a historical welding process as a historical welding image, establishing an image training set, and constructing a frequency domain feature dissociation model; S32, defining a combined loss function, guiding the frequency domain feature dissociation model to characterize the frequency domain features, and performing supervised learning on the frequency domain encoding and dissociation results; S33, using the frequency domain feature dissociation model to extract different frequency domain codes and reconstruction codes in the optical tomography image collected in real time, and removing invalid features and redundant features through feature selection and dimensionality reduction processing, so as to dissociate and obtain feature information that meets the requirements of power battery welding detection; The feature information includes low-frequency features, medium-frequency features and high-frequency features; wherein the low-frequency features represent the overall morphology and large-scale structural information of the weld, the medium-frequency features represent the medium-scale structural information of the weld, and the high-frequency features represent the surface detail information and edge information of the weld.

2. The power battery laser welding detection method integrating photoelectric-OCT in-situ sensing according to claim 1 is characterized in that: The method of obtaining the position and real-time posture information of the laser welding head, combining the surface structure image in the optical tomography image, tracking the welding seam trajectory in real time, and dynamically adjusting the welding position of the laser welding head comprises the following steps: S21, establishing a welding space coordinate of a laser welding system, and obtaining a coordinate position and a real-time posture of a laser welding head in the welding space coordinate; S22, generating a surface structure image of the optical tomography image by synthesizing multiple frames of the optical tomography image, and performing preprocessing and edge detection on the surface structure image to obtain an edge image to be welded; S23, extracting ideal weld feature points by fitting the edge line to be welded in the edge image to be welded, and forming an ideal weld trajectory; S24, tracking the ideal weld trajectory, calculating the position deviation between the laser welding head and the ideal weld in real time, and correcting the welding position of the laser welding head through feedback.

3. The power battery laser welding detection method integrating photoelectric-OCT in-situ sensing according to claim 2 is characterized in that: The method of generating a surface structure image of the optical tomography image by synthesizing multiple frames of the optical tomography image, and preprocessing and edge detecting the surface structure image to obtain the edge image to be welded comprises the following steps: S221, acquiring a plurality of frames of optical tomographic images captured continuously, aligning the plurality of frames of optical tomographic images using an image registration algorithm, and fusing and superimposing the images to form a surface structure image; S222, performing noise removal, grayscale equalization and image smoothing processing of the surface structure image in sequence; S223, performing a morphological opening operation on the surface structure image to eliminate sporadic pixel points in the background, and then performing a morphological closing operation to smooth the contour of the edge to be welded; S224, using an edge enhancement algorithm to perform edge enhancement on the surface structure image, then using an edge detection algorithm to extract the edge to be welded in the surface structure image, and using morphological operations to refine and correct the edge to be welded to obtain an image of the edge to be welded.

4. The power battery laser welding detection method integrating photoelectric-OCT in-situ sensing according to claim 2 is characterized in that: The step of extracting ideal weld feature points by fitting the edge line to be welded in the edge image to be welded to form an ideal weld trajectory comprises the following steps: S231, selecting the edge point of the edge image to be welded as the origin and establishing the image coordinates; S232, establishing a B-spline curve by setting control points and node vectors, and adjusting the B-spline curve by using the least square method to obtain a continuous and smooth edge line to be welded; S233, setting ideal weld feature points in the edge line to be welded, wherein the ideal weld feature points include a starting point, an end point, a turning point, a curvature change point and an equidistant point; S234, by changing the curve parameter values, searching for coordinate points that meet the ideal weld feature points on the edge line to be welded, and obtaining the pixel coordinates of the corresponding coordinate points to form an ideal weld trajectory.

5. The power battery laser welding detection method integrating photoelectric-OCT in-situ sensing according to claim 4 is characterized in that: The method of tracking the ideal weld trajectory, calculating the position deviation between the laser welding head and the ideal weld in real time, and correcting the welding position of the laser welding head by feedback comprises the following steps: S241, converting the pixel coordinates of the ideal weld trajectory into three-dimensional coordinates in the welding space coordinates by a hand-eye calibration method, and establishing an association relationship between the laser welding head and the weld trajectory; S242, using a cubic non-uniform rational B-spline method to calculate the corresponding position of the laser welding head on the ideal weld trajectory, and finding the weld point closest to the laser welding head on the ideal weld trajectory; S243, obtaining the position deviation between the laser welding head and the ideal weld trajectory by calculating the position difference between the laser welding head and the closest weld point, and correcting the position deviation of the laser welding head by using a controller.

6. The power battery laser welding detection method integrating photoelectric-OCT in-situ sensing according to claim 5 is characterized in that: The construction of the frequency domain feature dissociation model comprises the following steps: S311, construct a frequency domain dissociation encoder to extract frequency domain features of different levels from historical welding images and dissociate them into different frequency domain codes; S312, constructing a frequency domain dissociation decoder to decode different frequency domain codes and generate welding fault images of different frequency domain sub-bands corresponding to the historical welding images; S313, constructing a reconstruction decoder, generating a reconstruction code by performing feature fusion on different frequency domain codes, and then learning from the reconstruction code and generating a reconstructed image corresponding to the historical welding image.

7. The power battery laser welding detection method integrating photoelectric-OCT in-situ sensing according to claim 6 is characterized in that: The expression of the combined loss function is: Where L represents the combined loss function; L dwt represents the discrete wavelet change loss function; L r represents the regularization term loss function; Y represents the frequency domain sub-band image set formed in the historical welding image; represents a set of welding tomographic images separated in frequency domain; x represents the input optical tomographic image; y f Represents the frequency domain subband image corresponding to each frequency domain feature; represents the reconstructed image generated by the reconstruction encoder; f represents the frequency domain features extracted by the frequency domain dissociation encoder; It indicates that the welding fault images corresponding to different frequency domain features are generated by the frequency domain dissociation decoder; λ represents the regularization coefficient.

8. The power battery laser welding detection method integrating photoelectric-OCT in-situ sensing according to claim 6 is characterized in that: The method of selecting feature information as input variables, constructing a back propagation neural network model, and using the back propagation neural network model to perform defect detection and classification identification on the finished welding product of the power battery includes the following steps: S41, defining the network architecture and model parameters of the back propagation neural network model, and defining feature labels based on the type of feature information as model input variables; S42, using the leave-one-out method to divide the feature information into a training data set and a test data set, and performing model training and model tuning on the back propagation neural network model; S43, inputting the characteristic information of the power battery obtained by real-time analysis into the back propagation neural network model, detecting defects existing in the welding process of the power battery, and identifying the defect type.

Images