Battery FPC defect detection method and system based on laser imaging

Abstract

The invention relates to the technical field of defect detection, and discloses a battery FPC defect detection method and system based on laser imaging. The method comprises the following steps: carrying out full-coverage scanning on the FPC surface of the battery, and collecting a first surface laser reflection image and a second surface laser reflection image; performing environment interference compensation processing to obtain a first surface scattering signal image and a second surface scattering signal image; performing multi-scale feature extraction to obtain first feature data and second feature data; analyzing a light intensity change relation of adjacent areas to obtain first surface defect distribution correlation information and second surface defect distribution correlation information; and performing double-angle light intensity change feature classification identification on the first surface defect distribution associated information and the second surface defect distribution associated information, and outputting a comprehensive defect detection result. According to the method, the blind area of single-angle imaging is effectively eliminated, the comprehensiveness and reliability of defect detection are remarkably improved, and high-precision positioning and classification of defects are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of defect detection, and in particular to a battery FPC defect detection method and system based on laser imaging. Background Art

[0002] Flexible printed circuits (FPCs) are an indispensable component in battery manufacturing, and their quality directly affects the reliability and service life of the battery. Traditional FPC defect detection mainly relies on manual visual inspection or single-angle optical imaging technology, which faces problems such as insufficient accuracy, low efficiency and poor stability. Especially in complex industrial production environments, factors such as ambient light changes and FPC surface reflections cause unstable image signal waveforms, which seriously affect detection performance. For tiny defects such as pinholes with a diameter of less than 20μm and open circuits with a width of less than 15μm, the detection capabilities of existing technologies are seriously insufficient and cannot meet the quality requirements of high-precision battery production.

[0003] Laser imaging technology, with its high directivity and focus, provides new possibilities for FPC defect detection, but laser scanning at a single angle has obvious blind spots. Due to the complexity of the FPC surface structure, some defects may be difficult to detect at certain angles due to the light scattering characteristics. At the same time, most of the current image processing algorithms are designed for static images and lack effective use of time-series change characteristics, making it difficult to accurately track and identify the complete trajectory of defects in a continuous scanning sequence. Summary of the invention

[0004] The main purpose of the present invention is to provide a battery FPC defect detection method and system based on laser imaging. The present invention effectively eliminates the blind spots of single-angle imaging, significantly improves the comprehensiveness and reliability of defect detection, and realizes high-precision positioning and classification of defects.

[0005] To achieve the above object, the present invention provides a battery FPC defect detection method based on laser imaging, comprising the following steps: Perform full coverage scanning on the battery FPC surface, and collect the laser reflection image of the first surface and the laser reflection image of the second surface; Performing environmental interference compensation processing on the first surface laser reflection image and the second surface laser reflection image to obtain a first surface scattering signal image and a second surface scattering signal image; Performing multi-scale feature extraction on the first surface scattering signal image and the second surface scattering signal image to obtain first feature data and second feature data; Analyze the light intensity variation relationship of adjacent areas in the first scanning direction according to the first characteristic data to obtain first surface defect distribution association information, and analyze the light intensity variation relationship of adjacent areas in the second scanning direction according to the second characteristic data to obtain second surface defect distribution association information; The first surface defect distribution associated information and the second surface defect distribution associated information are subjected to dual-angle light intensity change feature classification and identification, and a comprehensive defect detection result is output.

[0006] The present invention also provides a battery FPC defect detection system based on laser imaging, comprising: A scanning module is used to perform full coverage scanning on the surface of the battery FPC and collect the laser reflection image of the first surface and the laser reflection image of the second surface; A compensation module, used for performing environmental interference compensation processing on the first surface laser reflection image and the second surface laser reflection image to obtain a first surface scattered signal image and a second surface scattered signal image; A feature extraction module, used for performing multi-scale feature extraction on the first surface scattering signal image and the second surface scattering signal image to obtain first feature data and second feature data; an analysis module, configured to analyze the light intensity variation relationship of adjacent areas in a first scanning direction according to the first characteristic data to obtain first surface defect distribution association information, and to analyze the light intensity variation relationship of adjacent areas in a second scanning direction according to the second characteristic data to obtain second surface defect distribution association information; A classification and recognition module is used to classify and recognize the dual-angle light intensity change characteristics of the first surface defect distribution association information and the second surface defect distribution association information, and output a comprehensive defect detection result.

[0007] In summary, the technical solution provided by the present invention adopts dual-angle laser polarization scanning technology to fully scan the FPC surface through two vertically set incident angles, effectively eliminating the blind spots of single-angle imaging, and significantly improving the comprehensiveness and reliability of defect detection. The baseline autoencoder environmental compensation technology can adaptively extract and reconstruct baseline signals for different lighting environments, effectively suppress the interference caused by factors such as ambient light and surface reflection, and enable the detection system to maintain stable performance under light changes. The time-aware pyramid network effectively retains the characterization information of tiny defects through multi-scale feature extraction and spatial attention mechanism, especially for pinholes with a diameter of less than 20μm and open circuit defects with a width of less than 15μm. It has enhanced detection capabilities. The edge regularized jump graph structure solves the problem of over-smoothing through sparse constraints and long-distance connection mechanisms, can capture defect change information on a long time scale, and provide a reliable graph structure representation for defect trajectory tracking. The cross-validation fusion processing technology combines the defect distribution information in two orthogonal directions, and realizes globally optimized defect trajectory tracking through the minimum cost flow algorithm, which effectively improves the detection capability of tiny defects and low-contrast defects. The defect classification technology based on the dual-angle light intensity change characteristic pattern can accurately distinguish four types of defects: open circuit, pinhole, copper foil missing and foreign matter, and achieve high-precision positioning and classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 It is a schematic diagram of the steps of a battery FPC defect detection method based on laser imaging in one embodiment of the present invention; Figure 2 It is a structural block diagram of a battery FPC defect detection system based on laser imaging in one embodiment of the present invention.

[0009] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0010] 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.

[0011] Reference Figure 1 , this embodiment provides a battery FPC defect detection method based on laser imaging, comprising the following steps: S1, perform full coverage scanning on the battery FPC surface, and collect the laser reflection image of the first surface and the laser reflection image of the second surface; Among them, the first laser polarization light source is installed on one side of the X-axis path of the scanning platform, and its spatial position is fixed by the optical path stabilization structure inside the system to ensure that the laser beam always irradiates the copper foil surface of the battery FPC at a set incident angle. The wavelength of the first laser light source is set to 532nm to enhance the penetration and reflection sensitivity of the microstructure on the copper foil surface, and the polarization direction of its output light is controlled by a high-precision polarizer, so that it has the optical characteristics of consistent direction and single polarization state, thereby improving the structural resolution of the reflected signal. In the actual scanning process, the first incident angle is set to 30°, that is, the laser beam presents a specific tilt angle relative to the surface of the FPC copper foil layer to form a reflected image with directional sensitivity. At the same time, a second laser polarization light source is installed in the orthogonal direction of the platform, which is arranged vertically in space with the first light source, so that the laser beam generated by the light source scans the FPC along the Y direction, and the incident angle is set to 45° perpendicular to the first incident direction. This angle setting constitutes an orthogonal system of scanning directions and enhances the perception of different microstructure orientations. The second laser light source also uses a polarized laser with a wavelength of 532nm, and adjusts its incident polarization state and direction through an independent polarizer control system. During the entire scanning process, the two laser sources are started in turn, the first light source performs X-direction scanning, and the second light source completes Y-direction scanning, and cooperates with the linear array CCD image sensor to collect the laser reflection intensity on the scanning track in real time. The system uses a synchronous controller to accurately bind the sampling timing to the position encoder data of the scanning platform to ensure that each sampling point corresponds to the accurate coordinate position of the FPC surface to avoid offset caused by mechanical errors or laser speckles. During the X-direction laser scanning process, the acquisition system continuously records the laser light intensity signal reflected from the FPC copper foil layer to form the original X-direction light intensity data; the original Y-direction light intensity data is also collected during Y-direction scanning. These light intensity data correspond to different sampling points in the two-dimensional space of the FPC surface in the form of a one-dimensional sequence, and are reorganized into a two-dimensional grayscale image through an image reconstruction algorithm. The image reconstruction process is based on scanning trajectory mapping and interpolation reconstruction technology, projecting the sampling point data into an equally spaced image grid, and using linear interpolation or bicubic interpolation methods to complete the unsampled areas in the image to ensure image continuity and smoothness. After the reconstruction is completed, the original X-direction light intensity data is reconstructed into the first surface laser reflection image, and the original Y-direction light intensity data is reconstructed into the second surface laser reflection image.

[0012] S2, performing environmental interference compensation processing on the first surface laser reflection image and the second surface laser reflection image to obtain a first surface scattered signal image and a second surface scattered signal image; Specifically, the first surface laser reflection image is input into the path fusion module of the baseline autoencoder. The path fusion module integrates the light intensity distribution of multiple sampling segments in the same scanning path to establish a joint feature expression reflecting the local environmental illumination state, surface reflection trend and system noise background, and generates a first surface environment feature vector with discriminative ability. The vector represents the fusion of multiple dimensions of information such as illumination non-uniformity, surface microstructure reflection anomaly and imaging system noise distribution. Similarly, after the second surface laser reflection image is input into the path fusion module, the second surface environment feature vector is obtained to describe the optical environment parameters under another laser incident angle. The first surface environment feature vector and the second surface environment feature vector are respectively input into the decoder part of the baseline autoencoder. The decoder is constructed with a symmetrical structure, which consists of five layers of deconvolution modules. Each layer contains deconvolution operations, batch normalization and nonlinear activation functions, and upsamples and restores the image structure layer by layer. In the first surface data processing, the decoder reconstructs a first surface baseline image without defects under an ideal state based on the first surface environmental feature vector. This image represents the theoretical imaging result that only contains environmental response and does not contain defect scattering under the current lighting conditions and reflection mode; in the second surface data processing, the decoder generates a second surface reconstructed baseline image, that is, an ideal reflection image under the assumption of no defects. The original laser reflection image of the first surface and the first surface reconstructed baseline image are subjected to pixel-by-pixel difference operation. This difference operation is based on the grayscale intensity of the pixel point, calculates the residual between the actual observed signal and the ideal baseline signal, and constitutes the first surface scattering signal image, which only retains abnormal structures that are inconsistent with the environmental model, that is, irregular scattering information that is considered to be caused by potential defects. Similarly, the original second surface laser reflection image and its corresponding reconstructed baseline image are subjected to difference processing to generate a second surface scattering signal image.

[0013] S3, performing multi-scale feature extraction on the first surface scattering signal image and the second surface scattering signal image to obtain first feature data and second feature data; It should be noted that the first surface scattering signal image is input into the first layer of the time-aware pyramid network, which uses a larger seven-by-seven convolution kernel, combined with a small step size operation and batch normalization processing, to extract the basic texture structure and low-frequency signal features in the image, and the convolution direction is set to the vertical direction to enhance the response to the longitudinal extension mode of the defect, and generate the first layer of vertical feature map; at the same time, the second surface scattering signal image is input into the first layer of the same network structure, and its convolution direction is adjusted to the horizontal direction, so that the network can recognize the basic texture regularity and small horizontal defect mode along the horizontal direction, and obtain the first layer of horizontal feature map. The first layer of vertical feature map is sequentially input into the second to fifth layers of the time-aware pyramid network, and each layer uses a standard convolution kernel of size three by three and a downsampling structure of step size two, so that each layer gradually expands the receptive field and extracts high-level semantic features. After the feature map of each layer is generated, the spatial attention mechanism is introduced. This mechanism gives higher response weights to local abnormal areas by constructing an attention weight matrix between pixels on the feature map, thereby highlighting the expression of small defects in the feature map, suppressing background redundant information, and forming a multi-scale vertical feature pyramid structure. Similarly, the first-layer horizontal feature map is input into the second to fifth layers of the convolution structure with the same configuration. After multiple layers of gradual abstraction and spatial attention weight enhancement, a multi-scale horizontal feature pyramid covering different semantic levels is generated. In order to integrate the detailed information and abstract features contained in the output of each layer, the feature pyramid structures of these two directions are fused through a horizontal connection mechanism. The horizontal connection includes the splicing and fusion convolution of the feature map channels, supplemented by normalization processing to ensure the consistency of the scale and amplitude of the feature maps at different levels. In the fusion process, the position encoding module is introduced. This module establishes implicit position perception capabilities in the time dimension by encoding the temporal position of each position in the feature map in the input sequence, so that the network can recognize the temporal evolution pattern of defects on the scanning path. The fused multi-scale vertical feature map is integrated into the first feature data after completing the integration and temporal encoding of all levels, representing the full-scale, high-semantic, and temporal-aware feature vector extracted based on the first surface scattering signal image; and the horizontally fused feature map forms the second feature data, which is used as the expression of the defect structure contained in the second surface scattering image.

[0014] S4, analyzing the light intensity variation relationship of adjacent areas in the first scanning direction according to the first characteristic data to obtain first surface defect distribution association information, and analyzing the light intensity variation relationship of adjacent areas in the second scanning direction according to the second characteristic data to obtain second surface defect distribution association information; Specifically, based on the first feature data, it is divided into multiple nodes according to the time series or scanning trajectory sequence, each node corresponds to a high-dimensional feature expression of a local area on the FPC surface, and then the cosine similarity calculation is performed on the feature vectors between all possible node pairs in the first scanning direction to construct a similarity matrix between node pairs, and the initial association graph of the first direction is generated based on the matrix. In this initial graph, all node pairs with similarities greater than a preset threshold (for example, 0.75) are connected by edges to form a graph structure with local similarity logic. The second feature data is processed in the same way, and the initial association graph of the second direction is generated by calculating the cosine similarity between each node pair in the second scanning direction, so that it expresses the structural and light intensity similarity relationship between the FPC surface areas in the Y direction. The initial graph structure has problems such as connection redundancy, local over-density and remote missing, and an edge regularization jump connection mechanism is introduced to enhance the expression ability of the graph. The mechanism sets the maximum number of edge connections K for each node to limit the number of adjacent edges and avoid feature weakening caused by oversmoothing. At the same time, in order to improve the modeling capability of defect trajectories across time domains, nodes are allowed to jump to nodes within the range of non-continuous time steps (such as ±5 steps before and after), so as to capture long-range dependencies and time-span defect propagation patterns. The edge regularization mechanism also introduces an edge cost evaluation model, which integrates three types of information: feature similarity, spatial proximity, and light intensity gradient between nodes, as the basis for edge existence and weight setting, thereby suppressing noise connections and highlighting real defect paths. After completing the construction of the regularized jump graph in the first and second directions, node embedding learning operations are performed on them respectively. The graph neural network used contains three layers of graph convolution modules, each layer is configured with a different number of convolution kernels (for example, 32, 64, 128). The representation capability of each node in its local neighborhood structure is extracted through feature aggregation and activation function processing, and the edge weights are dynamically allocated through the attention mechanism to achieve weighted aggregation of node features. After learning through the graph neural network, each node in the first-direction regularized jump graph carries its embedded expression in the local structure, thereby forming the first surface defect distribution association information. Similarly, the second-direction regularized jump graph obtains the second surface defect distribution association information after embedding learning.

[0015] S5, classifying and identifying the dual-angle light intensity change characteristics of the first surface defect distribution association information and the second surface defect distribution association information, and outputting a comprehensive defect detection result.

[0016] Among them, a spatial correspondence is established between the first surface defect distribution association information and the second surface defect distribution association information. A registration algorithm based on node position encoding and scanning sequence remapping is used to normalize and match the coordinates of the graph node sets in the two directions, so that the defect features of the same physical position can be structurally aligned in two perspectives, and the registered bidirectional defect distribution data is generated. Based on the registered bidirectional defect distribution data, the probability of defect existence is calculated for each detection position. By estimating the probability of defect classification of the embedded features of each node in the first scanning direction and the second scanning direction, the probability of independent defect existence of the position in two perspectives is obtained respectively. The defect probabilities in the two directions are normalized using a weighted fusion strategy to construct an initial fused defect probability graph. Based on the fused probability graph, a defect trajectory candidate set is established according to the high-probability response nodes in the graph. By setting all nodes that may be the starting points of the defect trajectory as source points and all possible terminal nodes as sinks, a global connection structure covering all potential defect paths is established on the entire fusion graph. On this basis, in order to screen out the optimal defect trajectory expression, a global optimization mechanism based on the minimum cost flow model is introduced. For each candidate trajectory path, its appearance cost (reflecting whether the node looks like a defect), transfer cost (reflecting the rationality of forming a trajectory between adjacent nodes) and missing cost (corresponding to the penalty for missing nodes or low-confidence nodes in the trajectory) are calculated respectively, and these are used as the three dimensions of the trajectory cost function. By solving the minimum cost flow problem, a set of optimized defect trajectory sets covering the entire defect area is obtained. The trajectory splicing operation is performed on the optimized trajectory, and multiple trajectories that may cross regions but have continuous structures are combined into a complete defect band structure to construct a surface defect distribution network covering the entire FPC surface. Based on this network, the light intensity change sequence on each trajectory path is extracted, the reflection mode and change trend of the defect along the trajectory are analyzed, and a three-dimensional structure diagram is generated in combination with the node spatial depth position information to construct a surface three-dimensional defect distribution map with three-dimensional spatial expression capabilities. In this three-dimensional map, each defect instance contains not only its two-dimensional projection coordinates, but also multi-dimensional data such as its light intensity profile, change frequency, and morphological contour. Based on this three-dimensional expression, the defect pattern is mapped to a predefined set of light intensity change templates. Through template matching and supporting neighbor classification technology, four typical defects, namely, open circuit (manifested as long strips of continuous low light intensity), pinholes (manifested as isolated high reflection points), copper foil missing (manifested as regional high light intensity reflection) and foreign matter (manifested as irregular fluctuations), are classified and identified, and the final comprehensive defect detection results are output, which include the defect location coordinates, category type, size range, severity level and confidence probability score.

[0017] The time-series light intensity signal sequence of each significant defect area is extracted from the surface three-dimensional defect distribution map. The sequence originates from the light intensity change record in the graph structure trajectory formed by the bidirectional scanning trajectory, reflecting the light reflection response curve of the defect in the two laser incident directions. The light intensity sequence is input into the light intensity change pattern analysis module. By analyzing the peak value, valley value, change rate, continuous time, symmetry and other dimensions of the light intensity, a dual-angle defect feature parameter set describing the spatiotemporal evolution characteristics of each defect is constructed, which contains the horizontal and vertical light intensity evolution path characteristics and integrates its spatial distribution and time jump characteristics. Based on the above dual-angle feature parameter set, the nearest support neighbor analysis technology is used to compare and locate each defect area. The current defect parameter set is measured by multi-dimensional distance with the defect category features marked in the known sample library, and the most similar set of support neighbors is identified. The center distribution and defect probability distribution of this group of samples are used as the precise position estimation result of the current defect. In order to improve the defect positioning accuracy, the above defect precise position data is interpolated at the sub-pixel level. Using quadratic surface fitting or high-order interpolation algorithm, the grayscale distribution of the defect center position within the sub-pixel range is approximated in the current image coordinate system, so that the positioning error of the high-precision defect positioning result is controlled within 5μm, which effectively meets the defect positioning requirements under high-density wiring. A set of light intensity mode feature thresholds are set for the dual-angle defect feature parameter set. These thresholds are set according to the typical light intensity change law of various defects. For example, the open circuit is manifested as a decrease in light intensity in the continuous area, the pinhole is manifested as a sudden change in high light intensity at an isolated point, the copper foil missing area is manifested as high-intensity diffuse reflection, and the foreign body defect usually presents irregular fluctuations. By matching the light intensity characteristics of the current defect with various feature templates, the defect type classification is achieved and the corresponding defect type identification is generated. The high-precision position result and type information of the defect are input into the contour fitting module. By performing boundary tracking and least squares fitting on the defect edge points, the complete defect contour shape is obtained, and the geometric feature parameters such as the area, aspect ratio, and morphological complexity of the defect are extracted to form the geometric feature data of the defect. The above-generated defect precise location data, defect type identification and defect geometric feature data are structured and integrated to output a comprehensive defect detection result including spatial coordinates, classification information, size parameters and severity.

[0018] In one example, a full coverage scan is performed on the battery FPC surface to collect a first surface laser reflection image and a second surface laser reflection image, including: The first polarized laser light source is fixed on a high-precision scanning platform, and the copper foil layer on the surface of the battery FPC is scanned in the X direction at a first incident angle by adjusting the wavelength of the first polarized laser light source to obtain an X-direction scanning laser reflection signal; The second polarized laser light source is fixed on a high-precision scanning platform, and the copper foil layer on the surface of the battery FPC is scanned in the Y direction by adjusting the wavelength of the second polarized laser light source and setting the second incident angle to be perpendicular to the first incident angle, so as to obtain a Y direction scanning laser reflection signal; The light intensity of the laser reflection signal scanned in the X direction and the laser reflection signal scanned in the Y direction are collected to obtain the original light intensity data in the X direction and the original light intensity data in the Y direction; The original X-direction light intensity data is reconstructed to obtain a first surface laser reflection image, and the original Y-direction light intensity data is reconstructed to obtain a second surface laser reflection image.

[0019] In this example, the first laser polarization light source is fixed at a specified position of the high-precision scanning platform, and its incident direction is adjusted by a spatial calibration mechanism so that the laser beam is irradiated onto the surface of the FPC copper foil layer at an incident angle of 30°. The first laser light source uses a high-stability laser with a wavelength of 532nm, which is in the visible green light range and has good focusing ability and structural penetration, and is suitable for detecting small differences on the surface of metal microstructures. When the light source emits a laser beam, its output is polarized by a set of adjustable polarizers to adjust the polarization state, so that the output light is in a specific linear polarization state, thereby enhancing the response capability to the anisotropic reflection characteristics of the surface and ensuring that the reflected signal contains richer directional light intensity information. The high-precision scanning platform is equipped with a multi-axis electric control slide and a position encoder, which controls the translation of the FPC sample in the X-axis direction with micron-level resolution, so that the laser spot advances at a constant speed along the entire scanning trajectory, and the reflected light intensity is instantaneously collected at each discrete sampling point through the linear array CCD camera below. Since the spot diameter is usually in the range of several microns and the CCD has sub-pixel sampling accuracy, it is possible to fully record the laser intensity data reflected from each tiny area to form a one-dimensional X-direction laser reflection signal sequence. In order to obtain the Y-direction reflection information that is completely orthogonal to the X-direction, a second laser polarization light source is installed on the platform simultaneously. Its fixing method and laser parameters are basically the same as the first light source, but the incident direction is set to be perpendicular to the first incident angle, that is, set to 45°, so that its laser beam covers the entire copper foil surface along the Y-axis direction perpendicular to the X-axis. Similarly, the second light source adjusts the polarization state through its exclusive polarization control component to ensure that it has the same direction recognition ability and reflection structure sensitivity in the Y direction. During the Y-direction scanning process, the scanning platform starts another axial motion trajectory to push the FPC sample to move evenly along the Y direction. The laser beam maintains a constant angle of irradiation with the sample surface. At the same time, the linear array CCD continuously records the laser reflection intensity of each position point, forming a Y-direction scanning laser reflection signal sequence independent of the X-direction. During the scanning process in the above two directions, the system's built-in synchronization control module records the position information of each sampling point in the platform coordinate system, and identifies all reflection signals in the X and Y directions according to the time series and spatial position, providing an accurate spatial index basis for subsequent data reconstruction. The initial state of the light intensity data of the laser reflection signal is a one-dimensional intensity sequence, which is converted into a two-dimensional reflection image through the image reconstruction algorithm. The image reconstruction process maps the one-dimensional signal interpolation to the two-dimensional image grid based on the laser scanning trajectory and sample displacement data. Linear interpolation, bilinear interpolation or bicubic interpolation are used to complete the uncovered image areas between the scanning trajectories, and median filtering or Gaussian smoothing is used to enhance the image structure coherence. The mapping accuracy of the image coordinates is determined by the platform encoder control accuracy and the CCD sampling resolution, ensuring that the reconstructed image can spatially reflect the actual reflection intensity distribution of the sample surface.After reconstruction, the original reflection signal in the X direction is mapped to the first surface laser reflection image, which takes the transverse scanning path as the main axis and presents the directional reflection characteristics of the copper foil surface at an incident angle of 30°; while the reflection signal in the Y direction generates the second surface laser reflection image, which reflects the light reaction state at an incident angle of 45°. The two images complement each other in structure, revealing the difference in optical response of the FPC surface structure in two polarization directions. The first surface image is more sensitive to linear defects or structural changes extending in the transverse direction, while the second surface image is more suitable for capturing discontinuities, point defects or surface foreign bodies in the vertical structure. The combination of the two provides the system with multi-perspective structural feature input.

[0020] In one example, performing environmental interference compensation processing on the first surface laser reflection image and the second surface laser reflection image to obtain the first surface scattered signal image and the second surface scattered signal image includes: Inputting the laser reflection image of the first surface into the path fusion module in the baseline autoencoder to extract global environmental features, thereby obtaining a first surface environmental feature vector; The laser reflection image of the second surface is input into the path fusion module in the baseline autoencoder to extract global environmental features, so as to obtain the environmental feature vector of the second surface; Reconstructing a signal based on a first surface environment feature vector by using a decoder of a baseline autoencoder to obtain a first surface reconstructed baseline image; Reconstructing a signal based on a second surface environment feature vector by using a decoder of the baseline autoencoder to obtain a second surface reconstructed baseline image; Performing a differential operation on the first surface laser reflection image and the first surface reconstructed baseline image to obtain a first surface scattering signal image; A difference operation is performed on the second surface laser reflection image and the second surface reconstructed baseline image to obtain a second surface scattering signal image.

[0021] In this example, the first surface laser reflection image is input into the path fusion module of the baseline autoencoder to integrate the spatial features of the image on a macro scale. The path fusion module jointly models the continuous image segments in the entire scanning path. Its internal structure consists of multiple convolution-pooling units, which can extract the environmental reflection feature pattern in the path, including the distribution trend of laser irradiation intensity, the difference in surface material reflection, and the distribution structure of background clutter. By fusion modeling of multiple scanning paths in the entire image, the global features of the current sample imaging environment are captured from a large-scale statistical perspective, and a stable and discriminative first surface environmental feature vector is generated. This vector contains key information such as illumination, structural directionality, texture frequency, and reflection contrast in the current image, and can provide reliable priors for subsequent baseline reconstruction without relying on specific defect labels. The first surface environmental feature vector is input into the decoder part of the autoencoder. The decoder structure uses a five-layer deconvolution stack to restore the ideal image state when there is no defect in theory based on the extracted environmental features. To ensure the spatial consistency and structural fidelity of the reconstructed image, each layer of the decoder is equipped with normalization, nonlinear activation, and upsampling modules, and some low-level spatial information is retained through a jump connection mechanism. During the decoding process, the network restores the image data consisting only of normal reflection structures, that is, the baseline image without scattering anomalies, local light intensity mutations, and sudden noise points, and generates the first surface reconstruction baseline image. This image represents the "normal state" version of the first surface laser image under the current optical environment and system parameters, and is the network's optimal fit to the statistical structure of the defect-free area. The second surface laser reflection image is input into the path fusion module to extract the second surface environment feature vector complementary to the first direction. Since the second surface image originates from different incident angles and polarization directions, the reflection path, spot distribution, and material response it contains will also change, so it is independently modeled and features are extracted to avoid mismatch errors caused by different viewing angles. After obtaining the second environment feature vector, it is sent to the decoder to generate the second surface reconstruction baseline image, which is also a defect-free reference image without the influence of abnormal scattering, used to represent the standard reflection state under an ideal environment. The original reflection image and the reconstructed baseline image are differentially operated pixel by pixel. In the first surface, the first surface laser reflection image and the first surface reconstruction baseline image are differentially operated, and the grayscale difference or relative change rate is calculated for the pixel values ​​at the corresponding positions of the two. The essence of this operation is to remove the common characteristic signals in the normal area and only retain the part that the decoder cannot restore. This part is the structural anomaly that deviates from the normal reflection mode, that is, the possible area with defects such as pinholes, fractures, foreign matter, and microcracks. The difference result constitutes the first surface scattering signal image, whose image brightness is mainly concentrated around the defect position, and the background area is suppressed to a value close to zero, thereby achieving high contrast enhancement of the defect signal.The same difference process is applied to the second surface image, and pixel-level difference is performed between the second surface laser reflection image and its reconstructed baseline image to generate a second surface scattered signal image.

[0022] In one example, multi-scale feature extraction is performed on the first surface scattering signal image and the second surface scattering signal image to obtain first feature data and second feature data, including: Inputting the first surface scattering signal image into the first layer of the time-aware pyramid network to perform basic texture feature analysis, and obtaining a first-layer vertical feature map; The second surface scattering signal image is input into the time-aware pyramid network to perform basic texture feature analysis to obtain the first-layer horizontal feature map; The vertical feature map of the first layer is sequentially input into the second to fifth layers of the time-aware pyramid network, and the spatial attention mechanism is applied to the output feature map of each layer to enhance the representation of the defect area, thus obtaining a multi-scale vertical feature pyramid. The horizontal feature map of the first layer is sequentially input into the second to fifth layers of the time-aware pyramid network, and the spatial attention mechanism is applied to the output feature map of each layer to enhance the representation of the defect area, thus obtaining a multi-scale horizontal feature pyramid. The feature maps of different levels in the multi-scale vertical feature pyramid are fused through horizontal connections, and position encoding is added to capture the temporal dependency to obtain the first feature data; The feature maps of different levels in the multi-scale horizontal feature pyramid are fused through lateral connections, and position encoding is added to capture the temporal dependency to obtain the second feature data.

[0023] In this example, the first surface scattering signal image is input into the first layer of the time-aware pyramid network, which is processed by a seven-by-seven convolution kernel with a large receptive field to extract the basic texture features of the FPC surface scattering image while maintaining the integrity of the spatial information. Since the scattering image has removed most of the background redundant signals through the differential operation of the autoencoder, the first layer of convolution retains the texture disturbance features caused by surface microcracks, pinholes or local copper foil disconnection to the greatest extent. The convolution operation direction of the first layer is set to the vertical direction, and the convolution kernel maintains complete coverage of the structure in the X direction during the scanning process. This directional design helps to enhance the response capability to vertically arranged defects, and obtains the first layer of vertical direction feature map as the starting feature reference of the subsequent pyramid network. At the same time, the second surface scattering signal image is input into the parallel processing branch of the same network, and the first layer of convolution module with the same parameter structure is used to perform basic feature extraction on it. The execution direction of the convolution kernel on this path is horizontal, focusing on capturing the structural features extending along the Y direction, which is suitable for identifying lateral defect responses such as linear fractures, desoldering or foreign body occlusion along the circuit routing direction. This operation outputs the first-layer horizontal feature map, which together with the first-layer vertical feature map form two initial feature channels with orthogonal structural perception capabilities. The two first-layer feature maps are respectively input into the second to fifth layers of TAPN. These layers adopt a gradually decreasing convolution kernel size (usually three by three) and a downsampling mechanism to reduce the image resolution in exchange for a larger receptive field, thereby extracting global semantic features while maintaining the local structure of the space. On each layer of the output feature map, a spatial attention mechanism is introduced to calculate the response importance of each pixel in the feature map to the entire image, and by learning a trainable weight distribution matrix, a higher response value is given to key areas such as defect edges and morphological mutation points, thereby improving the sensitivity and robustness of the network to defect representation. The processing of the first-layer vertical feature map forms a multi-scale vertical feature pyramid, which retains a multi-level feature set from low-level local textures to high-level semantic contours in structure; while the processing of the first-layer horizontal feature map generates a multi-scale horizontal feature pyramid, both of which reflect the continuous expression of defects at different scales and directions. Horizontal connection fusion is performed on the above two pyramid structures respectively. The lateral connection operation aligns and fuses feature maps from different scales in a unified dimension through channel splicing and feature convolution, improves the efficiency of information transmission between upper and lower levels, and strengthens the ability to express tiny defects without dilution in deep networks. During the fusion process, a position encoding module is introduced to embed the spatial position in each feature map into a temporal label to establish the time order of the pixels inside the feature map in the FPC scanning path. This encoding method is implemented through sine and cosine position functions or learnable position embedding vectors. Its role is to introduce time dependency in the network feature representation, so that the model can recognize the dynamic pattern of defects evolving over time on the scanning path.After the position code is embedded, the fused vertical pyramid features are integrated into the first feature data, which is a set of high-dimensional feature representations with direction perception, multi-scale semantic fusion and temporal dependence, used to capture the occurrence, evolution and diffusion patterns of small defects on the battery FPC surface during the X-direction scanning process. Similarly, the fused horizontal pyramid features are converted into the second feature data, which focuses on capturing structural anomalies and local change trends during the Y-direction scanning process.

[0024] In one example, the light intensity variation relationship of adjacent areas in the first scanning direction is analyzed according to the first characteristic data to obtain the first surface defect distribution association information, and the light intensity variation relationship of adjacent areas in the second scanning direction is analyzed according to the second characteristic data to obtain the second surface defect distribution association information, including: Calculate the cosine similarity matrix between all node pairs in the first scanning direction based on the first feature data to obtain an initial association graph in the first direction; Calculate the cosine similarity matrix between all node pairs in the second scanning direction based on the second feature data to obtain an initial association graph in the second direction; Applying an edge regularized jump connection mechanism to the initial association graph in the first direction to obtain a first direction regularized jump graph, and applying an edge regularized jump connection mechanism to the initial association graph in the second direction to obtain a second direction regularized jump graph; Node embedding learning is performed on the first direction regularized skip graph to obtain first surface defect distribution association information, and node embedding learning is performed on the second direction regularized skip graph to obtain second surface defect distribution association information.

[0025] In this example, the first feature data is obtained from the front pyramid feature extraction network. The feature data is a unified embedding representation of multiple image blocks arranged along the first scanning direction (such as the X direction) in the multi-scale space, and each node represents a high-dimensional feature vector of a fixed scanning area on the FPC surface. Based on this node set, the cosine similarity between all node pairs is calculated in turn to measure the structural similarity or texture consistency between any two nodes. The cosine similarity is calculated by taking the dot product of the feature vectors of the two nodes and dividing it by the product of their modulus lengths. This indicator stably reflects whether the two regions have consistent light intensity change trends, structural direction characteristics, or defect edge textures in the high-dimensional feature space. By calculating all node pairs one by one, a symmetric similarity matrix is ​​generated, in which each matrix element represents the correlation strength between the two regions in the current direction. Based on the similarity matrix, the system performs sparse connection screening according to the similarity threshold (for example, 0.75), retains the edge connections between the node pairs with significantly high similarity, and constructs the initial association graph in the first direction. The graph is structurally an undirected graph or a directed weak graph, whose nodes are scanning areas and edges are candidate paths of potential defect trajectories formed by feature similarity. The cosine similarity matrix between all node pairs in the second scanning direction is calculated based on the second feature data. The feature vectors of all image blocks are arranged according to the Y-direction scanning path, and the cosine similarity matrix between all node pairs in the second direction is constructed in the above manner, and the initial association graph in the second direction is generated. This graph focuses on depicting the light intensity continuity and defect coupling mode between each region in the vertical scanning direction, which is an orthogonal supplement to the structural modeling in the first direction. In order to improve the expression accuracy and modeling ability of the graph structure, the edge regularization jump connection mechanism is applied on the basis of the initial association graphs in the two directions, thereby solving the two extreme problems existing in the initial graph structure: on the one hand, the connections between high-similar regions are too dense, resulting in excessive smoothness of the graph and information redundancy; on the other hand, the long-distance weak association paths are not connected, resulting in the inability to express the long-distance propagation of defects. Therefore, an upper limit K of connection is set for each node, and only the K neighbor nodes that are most similar to the current node are retained to form edge connections, thereby effectively controlling the local density of the graph and realizing edge regularization. At the same time, in order to make up for the limitation that ordinary graph structures can only connect adjacent areas, the system allows the establishment of cross-distance jump connections in the graph, that is, nodes are allowed to connect not only with nodes in adjacent time steps or spatial positions, but also with nodes within the range of ±T steps before and after. Jump connections can effectively capture the non-continuous but related patterns of defects in surface distribution, especially for defects such as slender fractures, ribbon-shaped foreign bodies or intermittent pinholes. It has great modeling value. After the fusion of edge regularization and jump mechanism, the first direction regularized jump graph and the second direction regularized jump graph are obtained respectively. These two graph structures reflect the directional clustering tendency of defect areas, the propagation continuity of defect paths and the potential coupling law between abnormal areas in topological distribution.In order to extract the potential semantic information and structural embedding features of each node in these graph structures, the graph neural network embedding learning is performed on the above two regularized graph structures respectively. Multi-layer graph convolutional network or graph attention network is used for node update and feature propagation. In the first direction regularized jump graph, all nodes use their directional feature vectors as initial input and receive information from their neighboring nodes during the graph convolution process. The feature information of neighboring nodes is weighted and converged to the current node through the convolution weight matrix and edge weight mechanism, so that each node contains information features from its local neighborhood and jump distant neighbors after updating; after multi-layer convolution iteration, the system obtains a node set expressed by high-dimensional embedding, which constitutes the first surface defect distribution association information, which includes the texture features of local defects, and also contains topological information such as the similarity propagation structure between it and the surrounding nodes, the direction of the potential defect trajectory, the strength of the jump structure and the degree of local aggregation. The same graph embedding learning process is applied to the second direction regularized jump graph to obtain the second surface defect distribution association information. Its embedding result expresses the propagation map, point connectivity, structural similarity propagation and abnormal jump law of defects on the second scanning path.

[0026] In one example, the first surface defect distribution association information and the second surface defect distribution association information are subjected to dual-angle light intensity change feature classification and identification, and a comprehensive defect detection result is output, including: Establishing a spatial correspondence between the first surface defect distribution association information and the second surface defect distribution association information to obtain registered bidirectional defect distribution data; Based on the registered bidirectional defect distribution data, the defect existence probability is calculated for each detection position, and the defect probabilities in the first scanning direction and the second scanning direction are fused to obtain an initial fused defect probability map; According to the initial fusion defect probability graph, the source point is set to connect all the starting defect nodes, and the sink point is set to connect all the ending defect nodes to obtain the defect trajectory candidate set; According to the defect trajectory candidate set, the appearance cost, transfer cost and missing cost are calculated for each defect trajectory and the optimal flow allocation is solved to obtain the optimized defect trajectory set; Perform trajectory splicing on the optimized defect trajectory set to obtain a surface defect distribution network, and perform light intensity change pattern analysis based on the surface defect distribution network to generate a surface stereo defect distribution map with three-dimensional spatial information; According to the surface three-dimensional defect distribution map, the open circuit, pinholes, copper foil missing and foreign matter on the surface of the battery FPC are classified and identified, and the comprehensive defect detection results are output.

[0027] In this example, based on the defect distribution association information of the first surface and the second surface, a spatial correspondence is established for the data in these two directions to achieve physical coordinate alignment and feature semantic pairing of multi-view data. Since the first feature data and the second feature data come from orthogonal scanning paths respectively, they are represented as node sets and edge weight connection networks in two different directions in the graph structure. Through the spatial registration algorithm, the nodes in the two directions are mapped to a unified physical coordinate system to ensure that all nodes representing the same physical position can be matched in subsequent analysis. The registration process is completed by combining the displacement encoding data of the scanning platform, the pixel coordinates in image reconstruction, and the position encoding in the feature space. At the same time, feature similarity is used as a constraint to optimize node pair matching to form a registered bidirectional defect distribution data structure. Each node contains embedded features and corresponding defect response indicators in two directions at the same time. After completing the bidirectional registration, the defect response intensity of each node in the first and second directions is statistically analyzed with the position of each node as the center point, and the probability of defect existence is calculated. The calculation of this probability value takes into account the abnormal feature weights aggregated by the node during the graph convolution process, the structural edge weight distribution between the node and the adjacent nodes, and the local light intensity disturbance amplitude in the original scattering signal map, and converts it into a probability measure in the [0,1] interval through a normalization strategy. The probability values ​​in the two directions are fused to generate an initial fused defect probability map containing spatial coordinates and fused defect probabilities. Each pixel or node in the map corresponds to an FPC scanning position, and its value represents the confidence of the defect signal observed at the current position, which is the basic probability distribution for subsequent trajectory modeling. According to the initial fused defect probability map, the source point is set to connect all the starting defect nodes and the sink point is set to connect all the ending defect nodes to obtain a candidate set of defect trajectories. The local maximum point is selected from all nodes as the possible starting position of the defect trajectory and set as the source point. At the same time, the node in the edge area or the area with an obvious trend of defect intensity decrease in the spatial distribution is selected as the ending position, that is, the sink point. Then, a set of paths from the source point to the sink point is constructed in the entire graph. Each node in the path must satisfy a certain light intensity similarity and a progressive trend of probability values ​​between adjacent nodes to generate a candidate set of trajectories covering the entire defect area. After the trajectory candidate set is generated, in order to select the path structure with the best structure and the strongest defect expression ability, the minimum cost flow optimization model is introduced to globally optimize the trajectory path. In this model, three types of costs are defined for each edge in each trajectory path: the first is the appearance cost, which is the inverse of the probability of whether the current node has defect characteristics, which is used to measure the credibility of the point as a defective node; the second is the transfer cost, which reflects the transition smoothness and graph structure rationality of two adjacent nodes in the trajectory path. This value is a function of the spatial distance, direction change amplitude and feature similarity between the two nodes; the third is the missing cost, which is used to punish the node missing, probability discontinuity or graph structure break in the trajectory path.The three types of costs are integrated to construct the objective function, and all paths are jointly solved through the minimum cost flow algorithm (such as augmented path method, linear programming or graph cut optimization, etc.), so as to extract a set of defect trajectory paths with the lowest global cost to form an optimized defect trajectory set. Trajectory splicing is performed on the optimized defect trajectory set, and the trajectories with cross, parallel or spatial adjacent relationships in all trajectory paths are connected and merged to form a defect propagation network. The trajectory splicing process is based on the three criteria of spatial adjacency, directional consistency and synchronization of light intensity changes, forming a surface defect distribution network composed of multiple continuous paths. In this network, each trajectory segment not only contains spatial coordinate information, but also embeds its time evolution trajectory, light intensity change curve and graph structure information. Based on the light intensity change pattern of each node in the trajectory network, the pattern analysis is performed to construct a three-dimensional defect distribution map. The three-dimensional coordinates of the map correspond to the spatial position of the FPC, the evolution of the laser reflection intensity and the time or scanning trajectory sequence. By modeling the dynamic response of the defect light intensity change, the dynamic characteristic pattern of the defect type is identified. For example, the open circuit is manifested as a sudden drop in light intensity in a continuous area, the pinhole defect is manifested as an isolated high-intensity transient response, the copper foil loss forms a generalized reflection area, and foreign matter is mostly manifested as irregular or multi-peak structural disturbance. The three-dimensional structure map not only provides the spatial distribution of defects, but also explicitly encodes the optical behavior sequence, thereby greatly improving the credibility of subsequent classification and identification. Defect classification operations are performed based on the light intensity pattern, structural morphology and response change speed of each trajectory area in the three-dimensional defect map. By constructing a light intensity pattern template library, introducing support for neighbor matching, combining threshold discrimination and morphological factor analysis, all defects are divided into four categories: open circuit, pinhole, copper foil loss and foreign matter, and each type of defect is annotated with its start and end position, extension range, severity and classification confidence. All identified defects are output in the form of comprehensive defect detection results, which include both a graphical defect distribution annotation map and a structured data report that lists the spatial coordinates, type, size, morphological indicators and identification confidence score of each defect.

[0028] In one example, based on the surface three-dimensional defect distribution map, the open circuit, pinhole, copper foil missing and foreign matter on the surface of the battery FPC are classified and identified, and the comprehensive defect detection results are output, including: Extract the time-series light intensity signal sequence of each defect area from the surface three-dimensional defect distribution map, and perform defect light intensity change pattern analysis to obtain a dual-angle defect feature parameter set; Based on the dual-angle defect feature parameter set, the nearest support neighbor analysis is performed on each defect area to obtain the precise defect location data; Perform sub-pixel interpolation on the precise defect location data to obtain high-precision defect location results; The light intensity mode characteristic threshold is set according to the dual-angle defect characteristic parameter set, and the defects are classified into open circuit, pinhole, copper foil missing and foreign matter to obtain the defect type identification; Perform contour fitting processing on the high-precision defect location results and defect type identification to obtain defect geometric feature data; Integrate the precise defect location data, defect type identification and defect geometric feature data to output comprehensive defect detection results.

[0029] In this example, the surface three-dimensional defect distribution map is used as input. The map integrates the spatial coordinate information from the X and Y dual scanning directions, the fusion probability data and the trajectory structure features. Each defect trajectory represents the spatial extension path and the reflected optical response of a real defect. The light intensity values ​​of the nodes on each trajectory path are resampled and arranged in time or scanning order to form a complete time-series light intensity signal sequence. This sequence contains the reflection intensity when the laser beam irradiates the area, and also implicitly encodes the optical perturbation pattern caused by the projection of the defect on the scanning path, such as the position of the light intensity mutation, the speed of decline or rise, the symmetry of the waveform, the peak width, the valley duration and other multi-dimensional optical parameters. For each defect area, the corresponding light intensity sequence is extracted from the stereogram in the X and Y directions respectively, and the defect light intensity change pattern analysis is performed. This process is completed by local extreme point detection, signal derivative calculation, Fourier transform and time domain feature extraction. By encoding the features of the two-way sequences respectively, the time-series response information of the defect is converted into a unified feature parameter set, that is, the dual-angle defect feature parameter set. The parameter set includes: maximum reflection value, minimum reflection value, intensity change amplitude, dominant change direction, drop speed, average slope, fluctuation frequency, structural symmetry, light intensity mean, variance and other indicators, and is expressed in a unified vector structure to form a standard defect pattern descriptor that supports classification and positioning. The nearest neighbor analysis (KNN analysis) is performed based on the defect feature parameter set to achieve accurate position estimation of each defect area. A set of standard defect pattern libraries are pre-built, which contains typical light intensity parameter templates for open circuits, pinholes, copper foil missing and foreign objects generated by expert annotations or historical data. In the matching stage, the feature vector of the current defect area is used as the query point, and the K nearest neighbor samples with Euclidean distance or Mahalanobis distance are searched in the standard template library. Based on the center coordinate distribution and type labels of these neighbor samples, the spatial center point of the current defect is reversely inferred, which is the precise location data of the defect. In order to improve the positioning accuracy, sub-pixel interpolation processing is introduced. By constructing a continuous function model of light intensity distribution in the adjacent grid of the defect area (such as quadratic surface fitting, bicubic interpolation, spline interpolation, etc.), the light intensity extreme point is corrected by sub-pixel interpolation, and the positioning error is compressed to below the single pixel resolution, so as to obtain high-precision defect positioning results. While achieving positioning, the dual-angle defect feature parameter set is used to perform light intensity mode threshold judgment and classification recognition on the current defect. This process sets a set of discrimination rules and threshold ranges based on the typical light intensity change mode of each defect in the dual-angle response: for example, the open circuit defect is manifested as the light intensity continues to drop to a low value and remain stable within a time window, the pinhole defect is manifested as a short-period sudden peak at a certain time point, the copper foil missing defect shows a large area enhancement of light intensity in both directions, and the foreign body defect mostly presents irregular oscillation or local peak-valley alternating structure.Based on these pattern features, the system compares the light intensity sequence of the current defect with the preset classification rules, and performs decision trees, support vector machines or simple logical judgments to distinguish, and finally outputs the type identification of the defect, which not only indicates its category, but also comes with classification confidence or category probability distribution. The high-precision positioning results and defect type identification are used as input to perform defect contour fitting processing to generate geometric feature data of the defect. The contour fitting process adopts a method based on combining regional growth and edge extraction: directional regional expansion is performed around the center point of the defect, boundary points with obvious changes in light intensity values ​​are screened, and preliminary boundaries are generated in combination with reflection intensity gradient maps or structural similarity distribution maps; then polygon fitting, B-spline fitting or least squares ellipse fitting methods are used to smooth the boundary points and perform function modeling, and output closed contour curves. The system then calculates the geometric features of the defect such as area, main axis direction, aspect ratio, morphological complexity, and concavity based on the contour structure, and encodes them into standardized geometric parameter vectors to form a complete defect morphological expression. The precise location data of the defect, the defect type identification and the defect geometric feature data are structured and integrated to generate the final comprehensive defect detection result output.

[0030] Reference Figure 2 , this embodiment provides a battery FPC defect detection system based on laser imaging, including: Scanning module 1, used to perform full coverage scanning on the battery FPC surface, and collect the laser reflection image of the first surface and the laser reflection image of the second surface; Compensation module 2, used for performing environmental interference compensation processing on the first surface laser reflection image and the second surface laser reflection image to obtain the first surface scattered signal image and the second surface scattered signal image; A feature extraction module 3, used for performing multi-scale feature extraction on the first surface scattering signal image and the second surface scattering signal image to obtain first feature data and second feature data; Analysis module 4, used for analyzing the light intensity variation relationship of adjacent areas in the first scanning direction according to the first characteristic data to obtain the first surface defect distribution association information, and analyzing the light intensity variation relationship of adjacent areas in the second scanning direction according to the second characteristic data to obtain the second surface defect distribution association information; The classification and recognition module 5 is used to classify and recognize the dual-angle light intensity change characteristics of the first surface defect distribution association information and the second surface defect distribution association information, and output a comprehensive defect detection result.

[0031] In this embodiment, for the specific implementation of each unit in the above system embodiment, please refer to the above method embodiment, which will not be repeated here.

[0032] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, system, article or method including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, system, article or method. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the presence of other identical elements in the process, system, article or method including the element.

[0033] The above description is only a preferred embodiment of the present invention, and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A battery FPC defect detection method based on laser imaging, characterized in that: include: Perform full coverage scanning on the battery FPC surface, and collect the laser reflection image of the first surface and the laser reflection image of the second surface; Performing environmental interference compensation processing on the first surface laser reflection image and the second surface laser reflection image to obtain a first surface scattering signal image and a second surface scattering signal image; Performing multi-scale feature extraction on the first surface scattering signal image and the second surface scattering signal image to obtain first feature data and second feature data; Analyze the light intensity variation relationship of adjacent areas in the first scanning direction according to the first characteristic data to obtain first surface defect distribution association information, and analyze the light intensity variation relationship of adjacent areas in the second scanning direction according to the second characteristic data to obtain second surface defect distribution association information; The first surface defect distribution associated information and the second surface defect distribution associated information are subjected to dual-angle light intensity change feature classification and identification, and a comprehensive defect detection result is output.

2. The battery FPC defect detection method based on laser imaging according to claim 1 is characterized in that: The method of performing full coverage scanning on the battery FPC surface and collecting the first surface laser reflection image and the second surface laser reflection image comprises: Fixing a first polarized laser light source on a high-precision scanning platform, and scanning the copper foil layer on the surface of the battery FPC in the X direction at a first incident angle by adjusting the wavelength of the first polarized laser light source to obtain an X-direction scanning laser reflection signal; The second polarized laser light source is fixed on the high-precision scanning platform, and the wavelength of the second polarized laser light source is adjusted and set at a second incident angle perpendicular to the first incident angle, and the copper foil layer on the surface of the battery FPC is scanned in the Y direction to obtain a Y direction scanning laser reflection signal; Performing light intensity collection on the X-direction scanning laser reflection signal and the Y-direction scanning laser reflection signal to obtain original X-direction light intensity data and original Y-direction light intensity data; The original X-direction light intensity data is image reconstructed to obtain a first surface laser reflection image, and the original Y-direction light intensity data is image reconstructed to obtain a second surface laser reflection image.

3. The battery FPC defect detection method based on laser imaging according to claim 1 is characterized in that: The step of performing environmental interference compensation processing on the first surface laser reflection image and the second surface laser reflection image to obtain a first surface scattered signal image and a second surface scattered signal image includes: Inputting the first surface laser reflection image into a path fusion module in a baseline autoencoder to extract global environmental features, thereby obtaining a first surface environmental feature vector; Inputting the second surface laser reflection image into the path fusion module in the baseline autoencoder to extract global environmental features to obtain a second surface environmental feature vector; Reconstructing a signal based on the first surface environment feature vector by using a decoder of the baseline autoencoder to obtain a first surface reconstructed baseline image; Reconstructing a signal based on the second surface environment feature vector by using a decoder of the baseline autoencoder to obtain a second surface reconstructed baseline image; Performing a differential operation on the first surface laser reflection image and the first surface reconstructed baseline image to obtain a first surface scattering signal image; A difference operation is performed on the second surface laser reflection image and the second surface reconstructed baseline image to obtain a second surface scattering signal image.

4. The battery FPC defect detection method based on laser imaging according to claim 1 is characterized in that: The performing multi-scale feature extraction on the first surface scattering signal image and the second surface scattering signal image to obtain first feature data and second feature data includes: Inputting the first surface scattering signal image into the first layer of the time-aware pyramid network to perform basic texture feature analysis to obtain a first-layer vertical feature map; Inputting the second surface scattering signal image into the time-aware pyramid network to perform basic texture feature analysis to obtain a first-layer horizontal feature map; The first layer of vertical feature maps are sequentially input into the second to fifth layers of the time-aware pyramid network, and a spatial attention mechanism is applied to the output feature maps of each layer to enhance the representation of the defect area, so as to obtain a multi-scale vertical feature pyramid; The first layer of horizontal feature maps are sequentially input into the second to fifth layers of the time-aware pyramid network, and a spatial attention mechanism is applied to the output feature maps of each layer to enhance the representation of the defect area, so as to obtain a multi-scale horizontal feature pyramid; Fusing feature maps of different levels in the multi-scale vertical feature pyramid through lateral connections, and adding position encoding to capture temporal dependencies, to obtain first feature data; The feature maps of different levels in the multi-scale horizontal feature pyramid are fused through lateral connection, and position encoding is added to capture the temporal dependency to obtain the second feature data.

5. The battery FPC defect detection method based on laser imaging according to claim 1, characterized in that: The step of analyzing the light intensity variation relationship of adjacent areas in the first scanning direction according to the first characteristic data to obtain first surface defect distribution association information, and analyzing the light intensity variation relationship of adjacent areas in the second scanning direction according to the second characteristic data to obtain second surface defect distribution association information includes: Calculate the cosine similarity matrix between all node pairs in the first scanning direction based on the first feature data to obtain an initial association graph in the first direction; Calculate the cosine similarity matrix between all node pairs in the second scanning direction based on the second feature data to obtain an initial association graph in the second direction; Applying an edge regularized jump connection mechanism to the first direction initial association graph to obtain a first direction regularized jump graph, and applying an edge regularized jump connection mechanism to the second direction initial association graph to obtain a second direction regularized jump graph; Node embedding learning is performed on the first-direction regularized skip graph to obtain first surface defect distribution association information, and node embedding learning is performed on the second-direction regularized skip graph to obtain second surface defect distribution association information.

6. The battery FPC defect detection method based on laser imaging according to claim 1 is characterized in that: The step of performing dual-angle light intensity change feature classification and identification on the first surface defect distribution association information and the second surface defect distribution association information, and outputting a comprehensive defect detection result, includes: Establishing a spatial correspondence between the first surface defect distribution association information and the second surface defect distribution association information to obtain registered bidirectional defect distribution data; Based on the registered bidirectional defect distribution data, the defect existence probability is calculated for each detection position, and the defect probabilities in the first scanning direction and the second scanning direction are fused to obtain an initial fused defect probability map; According to the initial fusion defect probability graph, a source point is set to connect all starting defect nodes, and a sink point is set to connect all ending defect nodes to obtain a defect trajectory candidate set; According to the defect trajectory candidate set, the appearance cost, transfer cost and missing cost are calculated for each defect trajectory and the optimal flow allocation is solved to obtain an optimized defect trajectory set; Performing trajectory splicing on the optimized defect trajectory set to obtain a surface defect distribution network, and performing light intensity variation pattern analysis based on the surface defect distribution network to generate a surface stereoscopic defect distribution map with three-dimensional spatial information; According to the surface three-dimensional defect distribution map, the open circuit, pinhole, missing copper foil and foreign matter on the surface of the battery FPC are classified and identified, and the comprehensive defect detection result is output.

7. The battery FPC defect detection method based on laser imaging according to claim 6 is characterized in that: According to the surface three-dimensional defect distribution map, the disconnection, pinhole, missing copper foil and foreign matter on the surface of the battery FPC are classified and identified, and the comprehensive defect detection results are output, including: Extracting a time-series light intensity signal sequence of each defect area from the surface three-dimensional defect distribution map, and performing defect light intensity variation pattern analysis to obtain a dual-angle defect feature parameter set; Performing a nearest support neighbor analysis on each defect area based on the dual-angle defect feature parameter set to obtain accurate defect location data; Performing sub-pixel interpolation on the precise defect position data to obtain a high-precision defect positioning result; According to the dual-angle defect characteristic parameter set, the light intensity mode characteristic threshold is set, and the defects are classified into open circuit, pinhole, copper foil missing and foreign matter to obtain the defect type identification; Performing contour fitting processing on the high-precision defect positioning result and the defect type identification to obtain defect geometric feature data; The precise defect location data, the defect type identification and the defect geometric feature data are integrated to output a comprehensive defect detection result.

8. A battery FPC defect detection system based on laser imaging, characterized in that: The steps for implementing the battery FPC defect detection method based on laser imaging according to any one of claims 1 to 7, the battery FPC defect detection system based on laser imaging comprises: A scanning module is used to perform full coverage scanning on the surface of the battery FPC and collect the laser reflection image of the first surface and the laser reflection image of the second surface; A compensation module, used for performing environmental interference compensation processing on the first surface laser reflection image and the second surface laser reflection image to obtain a first surface scattered signal image and a second surface scattered signal image; A feature extraction module, used for performing multi-scale feature extraction on the first surface scattering signal image and the second surface scattering signal image to obtain first feature data and second feature data; an analysis module, configured to analyze the light intensity variation relationship of adjacent areas in a first scanning direction according to the first characteristic data to obtain first surface defect distribution association information, and to analyze the light intensity variation relationship of adjacent areas in a second scanning direction according to the second characteristic data to obtain second surface defect distribution association information; A classification and recognition module is used to classify and recognize the dual-angle light intensity change characteristics of the first surface defect distribution association information and the second surface defect distribution association information, and output a comprehensive defect detection result.

Images