Battery module defect detection method and system based on deep learning

Abstract

The embodiment of the application provides a battery module defect detection method and system based on deep learning, relates to the technical field of battery manufacturing quality detection services, generates a three-dimensional surface representation image by performing view angle correlation mapping processing on a surface image set of a battery module to be detected, calls a defect area pre-segmentation network to position a potential defect area of the three-dimensional surface representation image, generates a three-dimensional candidate defect area distribution graph with boundary contour coordinates, area volume parameters and internal texture feature vectors, analyzes defect features of each candidate defect area based on depth and texture dimension information, generates defect category judgment results and defect geometric form parameters, and finally generates a defect detection comprehensive report and outputs the defect detection comprehensive report to a production management terminal to trigger a defect product sorting operation. The application realizes full coverage, high-precision three-dimensional defect detection and automatic sorting of complex curved surface battery modules.

Description

Technical Field

[0001] This invention relates to the field of battery manufacturing quality inspection service technology, and more specifically, to a method and system for detecting battery module defects based on deep learning. Background Technology

[0002] With the rapid development of the new energy industry, battery modules, as core components of energy storage systems, directly affect the safety performance and lifespan of end products through their manufacturing quality. During the production process, various appearance defects may occur on the surface of battery modules, such as scratches, dents, bulges, color differences, and poor welds. Therefore, comprehensive and efficient appearance quality inspection of battery modules before they leave the factory is of significant industrial importance.

[0003] Currently, the detection of surface defects in battery modules in industrial production mainly relies on the following technical solutions: The first is the traditional machine vision inspection method based on single-view two-dimensional image acquisition. This method uses an industrial camera at a fixed station to acquire a planar image of the battery module surface from a single angle, and then uses image processing algorithms or shallow classifiers for defect identification. However, because battery modules usually have complex curved surface structures and polyhedral geometries, a single-view two-dimensional image can only reflect the appearance information of a local area, making it difficult to cover the entire outer surface of the module. Furthermore, there are serious information occlusion and viewpoint distortion problems in the transition areas of curved surfaces, resulting in a high rate of missed defect detection. The second is the inspection scheme based on multi-view image stitching. This scheme deploys multiple cameras at different stations to take pictures from multiple directions, and then stitches the images from each viewpoint together to obtain a more complete information about the module surface. However, this type of scheme usually treats each viewpoint image as independent planar data for processing, without establishing the geometric relationship between different views. The stitching accuracy in curved areas is insufficient, and the stitched image lacks depth dimension information, failing to accurately reflect the three-dimensional geometric shape of the defects. The third category is inspection solutions based on 3D reconstruction. These solutions use active measurement methods such as structured light or laser scanning to acquire point cloud data of the module surface, and then use point cloud analysis for defect identification. However, these solutions usually require point-by-point scanning, which is slow and cannot meet the real-time inspection requirements of high-speed production lines. Furthermore, the point cloud data itself lacks surface material texture information, which has significant limitations in distinguishing between surface color difference defects and geometric defects. Summary of the Invention

[0004] In view of this, the purpose of this invention is to provide a method and system for detecting battery module defects based on deep learning.

[0005] According to one aspect of the present invention, a battery module defect detection method based on deep learning is provided, the method comprising: A set of appearance images of the battery module to be tested is obtained by a preset multi-angle imaging device. The set of appearance images includes multiple battery module surface image units that are synchronously acquired from different imaging perspectives. Each battery module surface image unit has a view orientation mark and acquisition time mark corresponding to the imaging perspective. Viewpoint association mapping processing is performed on the set of appearance images. The surface image units of the battery module with different viewpoint orientation marks are projected onto a unified spherical mapping space. The viewpoint association correspondence between the surface image units of each battery module is established, and a three-dimensional surface characterization image of the battery module to be detected in the spherical mapping space is generated. The three-dimensional surface characterization image contains depth dimension information reflecting the curved surface structure of the battery module and texture dimension information reflecting the surface material properties of the battery module. A preset defect region pre-segmentation network is invoked to perform potential defect region localization processing on the three-dimensional surface characterization image, generating a three-dimensional candidate defect region distribution map containing several candidate defect regions. Each candidate defect region in the three-dimensional candidate defect region distribution map has boundary contour coordinate information, region volume parameters, and region internal texture feature vector. Based on the depth dimension information and the texture dimension information, multi-dimensional defect feature analysis processing is performed on each candidate defect region in the three-dimensional candidate defect region distribution map to generate the defect category determination result and defect geometric morphology parameters of each candidate defect region. Based on the defect category determination result and the defect geometric morphology parameters, a comprehensive defect detection report for the battery module to be tested is generated, and the comprehensive defect detection report is output to the production management terminal to trigger the defective product sorting operation.

[0006] According to another aspect of the present invention, a battery module defect detection system based on deep learning is provided, comprising: a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; the memory is used to store a computer program; and the processor is used to execute the computer program to implement the steps of the battery module defect detection method based on deep learning described above.

[0007] According to another aspect of the present invention, a readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, can perform the steps of the above-described deep learning-based battery module defect detection method.

[0008] Through any of the above methods, multiple surface image units of the battery module, each marked with a viewpoint orientation and acquisition time sequence, are simultaneously acquired using a pre-set multi-angle imaging device. Based on this, viewpoint association mapping processing projects each viewpoint image unit onto a unified spherical mapping space. This allows the complex curved surface structure of the battery module to be completely and distortion-free encoded into a spherical representation with depth dimension information, while retaining the texture dimension information reflecting surface material properties from the original images of each viewpoint. This achieves the simultaneous fusion of two heterogeneous information types—geometric depth and surface texture—in a single three-dimensional representation space, overcoming the inherent defects of existing technologies such as planar stitching schemes losing depth information and point cloud scanning schemes losing texture information. Furthermore, a defect region pre-segmentation network is invoked to perform potential defect analysis on the three-dimensional surface representation image. The initial localization of the defect area utilizes the network's ability to learn 3D spatial features to simultaneously acquire boundary contour coordinates, regional volume parameters, and internal texture feature vectors at the candidate defect area level. This enables subsequent multi-dimensional defect feature analysis to be performed on candidate areas with clear spatial ranges and rich feature descriptions, significantly reducing the computational overhead of exhaustive defect search in global 3D space and improving the accuracy of defect localization. Finally, a comprehensive inspection report is generated based on the defect category determination results and defect geometric morphology parameters, triggering sorting operations. This enables automated defect detection of battery modules with complex curved surface structures in a high-speed production line environment, achieving comprehensive coverage, high precision, high efficiency, and 3D defect morphology quantification capabilities, significantly improving the technical level of battery module quality control before shipment.

[0009] To make the above-mentioned objects, features and advantages of the embodiments of the present invention more apparent and understandable, a detailed description will be given below in conjunction with the embodiments and the accompanying drawings. Attached Figure Description

[0010] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 A schematic diagram of the components of the deep learning-based battery module defect detection system provided in an embodiment of the present invention is shown. Figure 2 The diagram shows a flowchart of the battery module defect detection method based on deep learning provided in an embodiment of the present invention. Detailed Implementation

[0012] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0013] The terms “first,” “second,” “third,” etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention described herein can be implemented, for example, in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0014] Figure 1 A schematic diagram of exemplary components of a deep learning-based battery module defect detection system 100 is shown. The deep learning-based battery module defect detection system 100 may include one or more processors 104, such as one or more central processing units (CPUs), each of which may implement one or more hardware threads. The deep learning-based battery module defect detection system 100 may also include any storage medium 106 for storing any kind of information such as code, settings, data, etc. Non-limitingly, for example, storage medium 106 may include any type of RAM, any type of ROM, flash memory device, hard disk, optical disk, etc. More generally, any storage medium can use any technology to store information. Further, any storage medium may provide volatile or non-volatile retention of information. Further, any storage medium may represent a fixed or removable component of the deep learning-based battery module defect detection system 100. In one case, when processor 104 executes dependent instructions stored in any storage medium or combination of storage media, the deep learning-based battery module defect detection system 100 may perform any operation of the associated instructions. The deep learning-based battery module defect detection system 100 also includes one or more drive units 108 for interacting with any storage medium, such as a hard disk drive unit, an optical disk drive unit, etc.

[0015] The deep learning-based battery module defect detection system 100 also includes input / output 110 (I / O) for receiving various inputs (via input unit 112) and providing various outputs (via output unit 114). A specific output mechanism may include a presentation device 116 and a dependent graphical user interface (GUI) 118. The deep learning-based battery module defect detection system 100 may also include one or more network interfaces 120 for exchanging data with other devices via one or more communication units 122. One or more communication buses 124 couple the components described above together.

[0016] The communication unit 122 can be implemented in any manner, such as via a local area network, a wide area network (e.g., the Internet), a point-to-point connection, or any combination thereof. The communication unit 122 may include any combination of hardwired links, wireless links, routers, gateway functions, and systems such as the Deep Learning-Based Battery Module Defect Detection System 100, governed by any protocol or combination of protocols.

[0017] Figure 2 The diagram illustrates a flowchart of a deep learning-based battery module defect detection method and system provided in an embodiment of the present invention. This deep learning-based battery module defect detection method and system can be derived from... Figure 1 The deep learning-based battery module defect detection system 100 shown in this embodiment is implemented using an automated battery module production line as an application scenario. At the end of the production line, a defect detection station integrating a multi-angle imaging device is deployed. This multi-angle imaging device consists of multiple industrial cameras arranged around the battery module to be inspected and multiple ring-shaped illumination sources. After the battery module to be inspected arrives at the inspection station via the conveyor belt, it triggers a position sensor, and the multi-angle imaging device simultaneously acquires surface images of the battery module from multiple perspectives. The acquired image data is transmitted to an edge computing device equipped with a graphics processing unit and a neural network inference acceleration chip. This edge computing device performs all subsequent processing steps, ultimately generating a comprehensive defect detection report and outputting it to the production management terminal. The production management terminal sends a rejection command to the sorting robot arm based on the inspection results.

[0018] Step S110: Obtain a set of appearance images of the battery module to be tested through a preset multi-angle imaging device. The set of appearance images includes multiple battery module surface image units that are synchronously acquired from different imaging perspectives. Each battery module surface image unit has a view orientation mark and acquisition time mark corresponding to the imaging perspective.

[0019] The edge computing device establishes a synchronous trigger connection with N industrial cameras in the multi-angle imaging device through a multi-channel image acquisition card. N is the total number of industrial cameras. Each industrial camera is fixed in a specific orientation in a spherical coordinate system with the geometric center of the battery module as the origin. Its optical center coordinates (CX, CY, CZ) and imaging optical axis direction vectors (DX, DY, DZ) have been accurately measured and stored in the calibration parameter file during the system calibration phase using the Zhang Zhengyou calibration method combined with a spherical target. When the battery module to be inspected is in place, the edge computing device simultaneously sends hardware trigger pulses to all industrial cameras. Each industrial camera exposes at the same time and generates a two-dimensional digital image matrix. The shape of each two-dimensional digital image matrix is ​​H×W×C, where H is the number of pixel rows in the image height direction, W is the number of pixel columns in the image width direction, and C is the number of color channels, C=1. After receiving image data from each industrial camera, the edge computing device adds two types of markers to each image: a view orientation marker that stores the azimuth angle θ and elevation angle φ of the industrial camera in the spherical coordinate system, and an acquisition timing marker that stores the synchronization trigger pulse sequence number corresponding to the frame image. All N images with the above markers constitute the appearance image set {I1, I2, ..., IN}, where the view orientation marker corresponding to the nth image IN is (θn, φn).

[0020] Step S120: Perform viewpoint association mapping processing on the appearance image set, project the battery module surface image units with different viewpoint orientation marks to a unified spherical mapping space, establish the viewpoint association correspondence between each battery module surface image unit, and generate a three-dimensional surface representation image of the battery module to be detected in the spherical mapping space. The three-dimensional surface representation image contains depth dimension information reflecting the curved surface structure of the battery module and texture dimension information reflecting the surface material properties of the battery module.

[0021] Step S121: Read the view orientation mark corresponding to each battery module surface image unit from the appearance image set one by one, and parse the optical center coordinates and imaging optical axis direction vector of the imaging camera in the preset spatial coordinate system corresponding to the view orientation mark.

[0022] The edge computing device retrieves the surface image unit In of the nth battery module and its viewing orientation marker (θn, φn) from the set of appearance images. It then searches a pre-built mapping table of viewing orientation and camera parameters based on (θn, φn). This mapping table is a key-value pair structure, where the key is the viewing orientation marker and the value is the calibration parameter set of the industrial camera corresponding to that orientation. The retrieved optical center coordinates (Cn, X, Cn, Y, Cn, Z) and imaging optical axis direction vector (Dn, X, Dn, Y, Dn, Z) of the industrial camera are obtained. The imaging optical axis direction vector is a normalized three-dimensional vector that satisfies (Dn, X)^2 + (Dn, Y)^2 + (Dn, Z)^2 = 1.

[0023] Step S122: Based on the optical center coordinates and the imaging optical axis direction vector, construct the ray inverse tracing mapping relationship between the pixel points of each battery module surface image unit and the spatial points on the preset spherical mapping space.

[0024] The preset spherical mapping space is a sphere with radius R, centered at the geometric center of the battery module. The value of R is greater than the radius of the circumscribed sphere of the battery module. For any pixel coordinate (u, v) on In, the edge computing device uses the intrinsic parameter matrix K of the industrial camera to back-project the pixel coordinate onto a ray direction in the camera coordinate system. The intrinsic parameter matrix K is a 3×3 matrix, where the elements (fx, fy, cx, cy) are the calibration focal length and principal point coordinates of the industrial camera. The back-projection calculation process is as follows: Construct a homogeneous vector p = [(u-cx) / fx, (v-cy) / fy, 1], and after normalizing its magnitude, obtain the ray direction vector rc in the camera coordinate system. Transform this direction vector to the world coordinate system using the camera extrinsic rotation matrix Rn to obtain the ray direction vector rw = Rn × rc in the world coordinate system. The intersection of a ray propagating along the rw direction from the optical center (Cn, X, Cn, Y, Cn, Z) with the sphere X^2 + Y^2 + Z^2 = R^2 is the corresponding spatial point Pn(u, v) of that pixel in the spherical mapping space. By traversing all pixels in In, a one-to-one mapping relationship is established between each pixel (u, v) in In and the spherical spatial point Pn(u, v). This mapping relationship is the ray tracing mapping relationship.

[0025] Step S123: Call the preset spatial coordinate transformation algorithm, based on the ray inverse tracing mapping relationship, to transform the two-dimensional pixel coordinates on the surface image unit of each battery module to the three-dimensional spherical coordinates in the preset spherical mapping space, and generate a spherical projection point cloud set corresponding to each surface image unit of the battery module.

[0026] For the nth battery module surface image unit In, the edge computing device traverses all valid pixel coordinates (u, v) in In and finds the three-dimensional coordinates (X, Y, Z) of the corresponding spherical space point Pn(u, v) through the mapping relationship established in step S122. The gray value I(u, v) of each pixel is combined with its spatial coordinates (X, Y, Z) to form a three-dimensional space point with gray value attribute. The set of all three-dimensional space points constitutes the spherical projection point cloud set Qn={(X, Y, Z, I)} corresponding to the nth battery module surface image unit.

[0027] Step S124: Perform point cloud fusion processing on the spherical projection point cloud set generated by all battery module surface image units to eliminate coordinate overlap and density differences between projection point clouds from different viewpoints, and generate the fused overall point cloud of the battery module surface.

[0028] The edge computing device merges all N spherical projection point cloud sets {Q1, Q2, ..., QN} into the same 3D space. For any two points in the merged point cloud whose spatial distance is less than the distance threshold dmerge, they are merged into a single spatial point. The 3D coordinates of the merged spatial point are taken as the arithmetic mean of the coordinates of the points before merging, and the grayscale value is taken as the arithmetic mean of the grayscale values ​​of the points before merging. For each spatial point in the point cloud, the number of points from different viewpoints within its neighborhood radius r is counted. If this number is less than the viewpoint coverage threshold Nmin, the point is marked as an isolated noise point and removed from the point cloud. After processing, the fused overall point cloud Qfused of the battery module surface is obtained.

[0029] Step S125: Extract the spatial coordinate distribution features of the overall point cloud on the surface of the battery module, perform spatial surface fitting processing on the overall point cloud on the surface of the battery module, and generate depth dimension information representing the surface structure of the battery module. The depth dimension information includes the normal direction information and radius of curvature information of each spatial point on the overall point cloud on the surface of the battery module.

[0030] For each spatial point in Qfused, a local plane fitting is performed on all spatial points within its neighborhood radius r. Principal component analysis (PCA) is used for fitting, and the covariance matrix of the neighborhood point coordinates is calculated. Eigenvalue decomposition is then performed on the covariance matrix, and the eigenvector corresponding to the smallest eigenvalue is the normal direction vector n = (nx, ny, nz) for that spatial point. Based on the projection distribution of neighborhood points along the normal direction, the radius of curvature ρ = 1 / (λmin / (λ1+λ2+λ3)) is calculated for that spatial point, where λ1≥λ2≥λ3 are the three eigenvalues ​​of the covariance matrix. The normal direction information and radius of curvature information of all spatial points together constitute the depth dimension information D = {(n, ρ)}.

[0031] Step S126: Perform texture primitive extraction processing on each battery module surface image unit to obtain a local texture primitive encoding set for each battery module surface image unit. Each texture primitive encoding in the local texture primitive encoding set represents the texture mode description information of a local pixel region.

[0032] For the nth battery module surface image unit In, the edge computing device uses the Local Binary Mode (LBSM) algorithm to extract texture primitive codes. For each pixel (u, v) in In, the gray values ​​of P equally spaced sampling points on the circumference of its surrounding area with a radius of rpixels are taken. The gray value of each sampling point is compared with the gray value of the center pixel. If it is greater than or equal to the gray value, it is encoded as 1; if it is less than the gray value, it is encoded as 0. The P comparison results are arranged clockwise to form a P-bit binary number, which is converted into a decimal integer to obtain the texture primitive code LBP(u, v) of the center pixel. Traversing all pixels in In, the local texture primitive code set Tn={LBP(u, v)} is obtained.

[0033] Step S127: Based on the viewpoint association correspondence, map the local texture primitive encoding set of each battery module surface image unit to the corresponding spatial point of the overall point cloud of the battery module surface to generate texture dimension information. The texture dimension information includes the texture primitive encoding and texture continuity coefficient of each spatial point on the overall point cloud of the battery module surface.

[0034] For each spatial point in Qfused, the corresponding pixel position of the spatial point under each viewpoint is found by reverse searching through the mapping relationship established in step S122. The corresponding local texture primitive code is extracted, and the mode of the texture primitive codes from multiple viewpoints is taken as the texture primitive code of the spatial point. For this spatial point, the difference between its texture primitive codes and those of each neighboring spatial point is calculated. The difference is defined as the Hamming distance in binary form of the two texture primitive codes, and the complement of the average difference is taken as the texture continuity coefficient. The texture primitive codes and texture continuity coefficients of all spatial points constitute the texture dimension information T={(LBP, c)}.

[0035] Step S128: Standardize the depth dimension information, normalize the texture dimension information, and perform data alignment and fusion processing on the processed depth dimension information and texture dimension information in a preset spherical mapping space to generate a three-dimensional surface representation image composed of the superposition of depth data layer and texture data layer.

[0036] The edge computing device performs Z-score normalization on the radius of curvature ρ of each spatial point in the depth dimension information. The normalization formula is ρ'=(ρ-μρ) / σρ, where μρ and σρ are the mean and standard deviation of all radii of curvature, respectively. It then performs Min-Max normalization on the texture primitive encoding of each spatial point in the texture dimension information. The normalization formula is LBP'=(LBP-LBPmin) / (LBPmax-LBPmin). The normalized and standardized data are aligned according to the spatial point coordinates, and the depth dimension and texture dimension are stored as two data layers of the 3D surface representation image: the depth data layer (Dlayer) stores the normal direction vector n and the normalized radius of curvature ρ' of each spatial point, and the texture data layer (Tlayer) stores the normalized texture primitive encoding LBP' and the texture continuity coefficient c of each spatial point. The two layers of data are aligned on the same spherical space grid to form the 3D surface representation image V.

[0037] Step S130: Call the preset defect region pre-segmentation network to perform potential defect region localization processing on the three-dimensional surface characterization image, and generate a three-dimensional candidate defect region distribution map containing several candidate defect regions. Each candidate defect region in the three-dimensional candidate defect region distribution map has boundary contour coordinate information, region volume parameters and region internal texture feature vector.

[0038] Step S131: Input the three-dimensional surface representation image into the multi-scale feature encoding module of the defect region pre-segmentation network. The multi-scale feature encoding module performs multi-scale convolution processing on the depth dimension information and texture dimension information of the three-dimensional surface representation image to generate depth multi-scale feature maps and texture multi-scale feature maps.

[0039] The multi-scale feature encoding module of the defect region pre-segmentation network includes parallel depth convolution branches and texture convolution branches, each employing a 3D convolutional neural network structure. The depth convolution branch receives the depth data layer Dlayer of the 3D surface representation image as input. The shape of Dlayer is Hv×Wv×Dv×Cdeep, where Hv, Wv, and Dv represent the number of voxels in the height, width, and depth directions after discretization of the spherical mapping space, respectively, and Cdeep is the number of input channels. Cdeep=4 corresponds to the three components of the normal direction and the normalized radius of curvature. The depth convolution branch consists of three cascaded 3D convolutional blocks. Each 3D convolutional block contains a 3D convolutional layer with a kernel size of K1×K1×K1, a 3D batch normalization layer, and a linear rectified activation function layer. The convolution stride in the first 3D convolutional block is S1, the second S2, and the third S3. S1, S2, and S3 are preset stride parameters, and spatial downsampling at different scales is achieved by adjusting the stride. Three 3D convolutional blocks output depth multi-scale feature maps {Fd1, Fd2, Fd3}, respectively. The texture convolution branch receives the texture data layer Tlayer of the 3D surface representation image as input. The shape of Tlayer is Hv×Wv×Dv×Ctex, where Ctex=2 corresponds to the normalized texture primitive encoding and texture continuity coefficient. The structure of the texture convolution branch is the same as that of the depth convolution branch. The three 3D convolutional blocks output texture multi-scale feature maps {Ft1, Ft2, Ft3}.

[0040] Step S132: Input the depth multi-scale feature map and the texture multi-scale feature map into the feature interaction fusion module of the defect region pre-segmentation network, and perform feature interaction fusion processing on the depth multi-scale feature map and the texture multi-scale feature map through the cross attention mechanism to generate the depth texture fusion feature volume.

[0041] The feature interaction fusion module performs cross-attention fusion on the depth feature map Fd and texture feature map Ft output at each scale level. Using Fd and Ft as input, Fd is first mapped to a query tensor Qd through a 1×1×1 convolutional layer, and Ft is mapped to a key tensor Kt and a value tensor Vt through another 1×1×1 convolutional layer. Qd, Kt, and Vt are all high-dimensional tensors of the same size. The attention weight matrix A = softmax((Qd·Kt) / (sqrt(dk))) is calculated, where dk is the channel dimension of the key tensor. The dot product operation is performed along the channel dimension, and softmax is performed along the spatial dimension. The attention weight matrix A is weighted and aggregated with the value tensor Vt to obtain the texture-to-depth enhancement feature Fd_enhanced = A·Vt. Symmetrically, Ft is mapped to the query tensor Qt, and Fd is mapped to the key tensor Kd and the value tensor Vd. Attention weights are calculated to obtain the depth-to-texture enhancement feature Ft_enhanced. Fd, Ft, Fd_enhanced, and Ft_enhanced are concatenated along the channel dimension, and then compressed to the same number of channels as the original feature map through a 1×1×1 convolutional layer, generating the fused feature at this scale level. The above operation is performed on each of the three scale levels to obtain fused feature maps at three scales. These fused feature maps are then restored to the same spatial size through a 3D upsampling layer and concatenated along the channel dimension. Finally, a 3×3×3 convolutional layer is used for cross-scale information integration, outputting the depth texture fusion feature volume Ffused.

[0042] Step S133: Input the deep texture fusion feature volume into the region candidate generation module of the defect region pre-segmentation network, and generate multiple three-dimensional candidate regions in the deep texture fusion feature volume through the three-dimensional region proposal algorithm. Each three-dimensional candidate region has anchor point coordinates, region size parameters and region boundary contour parameters.

[0043] The region candidate generation module sets multiple anchor boxes of preset sizes and ratios at each spatial location of the deep texture fusion feature volume Ffused. The anchor boxes are 3D cuboids, with preset sizes corresponding to small, medium, and large scales, and preset ratios including various combinations of length, width, and height ratios. For each anchor box, a fully convolutional sub-network predicts its confidence score as belonging to a defective foreground region, as well as its center coordinate offset (ΔX, ΔY, ΔZ) and size scaling factor (SW, SH, SD). Based on the confidence scores, the top K anchor boxes with the highest scores are selected. After offsetting and scaling their center coordinates and sizes, multiple 3D candidate regions {R1, R2, ..., RM} are obtained. Each 3D candidate region Rm has anchor coordinates (am, X, am, Y, am, Z), region size parameters (Wm, Hm, Dm), and region boundary contour parameters.

[0044] Step S134: Through the candidate region filtering layer of the defect region pre-segmentation network, multiple three-dimensional candidate regions are subjected to overlap filtering and confidence ranking processing to select several candidate defect regions that meet the preset overlap conditions and preset confidence conditions.

[0045] The candidate region filtering layer calculates the 3D intersection-union ratio (IU) between each pair of M 3D candidate regions. The IU is the ratio of the intersection volume to the union volume of two 3D cuboid regions. For candidate region pairs with an IU exceeding the overlap threshold τiou, the one with the higher confidence score is retained, and the one with the lower confidence score is discarded. After completing the overlap filtering, the remaining candidate regions are sorted from highest to lowest confidence score, and the K candidate regions with the highest confidence scores are selected as the filtered candidate defect regions.

[0046] Step S135: Perform region boundary correction processing on several candidate defect regions, use the three-dimensional boundary regression algorithm to adjust the boundary contour of each candidate defect region, and generate the boundary contour coordinate information and region volume parameters of each candidate defect region.

[0047] For each of the K selected candidate defect regions, a 3D boundary regression subnetwork is used for boundary correction. The input to the 3D boundary regression subnetwork is the local feature block in Ffused corresponding to the candidate defect region, and the output is the boundary correction parameters (ΔXc, ΔYc, ΔZc, ΔW, ΔH, ΔD) for the candidate defect region. The corrected center coordinates of the candidate defect region are the original center coordinates plus an offset, the corrected size is the original size multiplied by a scaling factor, and the corrected boundary contour is determined by the center coordinates and size. Based on the corrected boundary contour coordinates, the regional volume parameter Vk of the candidate defect region is calculated as: Vk = corrected W multiplied by corrected H multiplied by corrected D.

[0048] Step S136: Through the feature pooling module of the defect region pre-segmentation network, each candidate defect region is mapped onto the deep texture fusion feature volume, and the texture feature information in the corresponding spatial range of each candidate defect region is extracted to generate the region-internal texture feature vector of each candidate defect region.

[0049] The feature pooling module uses the 3D RoIAlign algorithm. For each candidate defect region, its corresponding spatial range on Ffused is uniformly divided into Gr×Gr×Gr grid sub-regions. Within each sub-region, a preset number of voxel values ​​are taken, and the feature value of each sampling point is obtained through trilinear interpolation. The feature values ​​of all sampling points within the sub-region are averaged to obtain the feature representation of the sub-region. The feature representations of all sub-regions are arranged in order and flattened into a one-dimensional vector. This one-dimensional vector is the internal texture feature vector fm of the candidate defect region, with dimensions of Gr×Gr×Gr×Cff, where Cff is the number of channels in Ffused.

[0050] Step S137: Generate a three-dimensional candidate defect region distribution map based on the boundary contour coordinates, region volume parameters, and internal texture feature vectors of each candidate defect region. The three-dimensional candidate defect region distribution map marks the position and contour of each candidate defect region in three-dimensional space.

[0051] The edge computing device aggregates the boundary contour coordinates, region volume parameters, and internal texture feature vectors of the K candidate defect regions. It then draws 3D bounding boxes in 3D space based on the center coordinates and boundary contour coordinates of each candidate defect region, and displays the results visually, forming a 3D candidate defect region distribution map. This map is stored as structured data, containing K records, each corresponding to a candidate defect region. The fields include boundary contour coordinates, region volume parameters, and internal texture feature vectors.

[0052] Step S140: Based on depth dimension information and texture dimension information, perform multi-dimensional defect feature analysis processing on each candidate defect region in the 3D candidate defect region distribution map to generate defect category determination results and defect geometric morphology parameters for each candidate defect region. The defect category determination results are generated by calling a preset defect category classifier.

[0053] Step S141: Extract the boundary contour coordinate information of the current candidate defect region from the three-dimensional candidate defect region distribution map, and extract the local depth information subset corresponding to the current candidate defect region from the depth dimension information based on the boundary contour coordinate information.

[0054] The edge computing device traverses K candidate defect regions in the 3D candidate defect region distribution map. For the k-th candidate defect region, it extracts the bounding box region determined by its boundary contour coordinate information. It then extracts the normal direction information and radius of curvature information within the bounding box region from the depth dimension information D to form a local depth information subset Dsub,k.

[0055] Step S142: Perform depth gradient field analysis on the local depth information subset, calculate the normal direction offset and gradient magnitude of each spatial point in the local depth information subset, and generate depth gradient distribution data that reflects the surface undulation characteristics of the current candidate defect region.

[0056] For each spatial point in Dsub,k, calculate the average of the dot product of its normal direction vector nk and the normal direction vectors of all spatial points in its neighborhood, and take the inverse cosine to obtain the normal direction offset Δα = arccos(avg(nk·nneighbor)). Calculate the gradient magnitude of the radius of curvature Gρ = |grad(ρ)|, where grad(ρ) is approximated by the central difference method. The three-dimensional central difference formula is grad(ρ) = [(ρ(i+1,j,l)-ρ(i-1,j,l)) / 2, (ρ(i,j+1,l)-ρ(i,j-1,l)) / 2, (ρ(i,j,l+1)-ρ(i,j,l-1)) / 2]. The (Δα, Gρ) of all spatial points constitute the depth gradient distribution data Gk.

[0057] Step S143: Based on the depth gradient distribution data, extract abnormal depth abrupt change regions in the current candidate defect region where the surface undulation degree exceeds the preset undulation threshold, and use the spatial distribution of abnormal depth abrupt change regions as a component of the defect geometric morphology parameters.

[0058] Edge computing devices mark spatial points in Gk where the normal direction offset Δα exceeds the normal offset threshold τα or the gradient magnitude Gρ exceeds the gradient threshold τg as anomalous depth mutation points. Connectivity component analysis is performed on these anomalous depth mutation points, aggregating spatially adjacent points into several anomalous depth mutation regions. For each anomalous depth mutation region, the major axis vector vlen, minor axis vector vwid, and normal depth vector vdep of its smallest circumscribed ellipsoid are calculated. These three vectors constitute the defect geometric morphology parameter Gparam, k={(vlen, vwid, vdep)}, where vlen indicates the length extension direction of the defect, vwid indicates the width expansion direction, and the magnitude of vdep indicates the degree of depth indentation.

[0059] Step S144: Extract the texture feature vector within the current candidate defect region from the 3D candidate defect region distribution map, and construct the texture feature distribution histogram of the current candidate defect region based on the texture feature vector within the region.

[0060] The edge computing device reads the texture feature vector fk inside the k-th candidate defect region, which has a dimension of Df. The value range of all elements in fk is divided into B equal-width intervals, and the proportion pfreq(b) of the number of elements falling into each interval b to the total dimension Df of the vector is counted. A texture feature distribution histogram Hk={pfreq(1), pfreq(2), ..., pfreq(B)} is constructed.

[0061] Step S145: Extract the texture primitive codes of the surrounding regions adjacent to the current candidate defect region from the texture dimension information, perform texture difference comparison processing on the texture primitive codes of the texture feature vector inside the region and the texture primitive codes of the surrounding regions, and generate texture difference feature parameters.

[0062] The edge computing device extracts all texture primitive codes from the texture dimension information T, covering the surrounding area within a width Δr extending outward from the boundary of the k-th candidate defect region. This constitutes the surrounding texture primitive code set LBPneighbor. The texture primitive code component LBPinside is separated from fk. The KL divergence between the statistical distributions of LBPinside and LBPneighbor is calculated: KL(LBPinside||LBPneighbor)=∑(pinside(x)×log(pinside(x) / pneighbor(x))), summing over all possible code values ​​x. pinside(x) and pneighbor(x) are the normalized coding frequencies of the interior and periphery, respectively. This KL divergence value is the texture difference feature parameter dtex,k.

[0063] Step S146: Normalize the texture feature vector, texture feature distribution histogram, and texture difference feature parameters within the region to generate a standardized texture descriptor; perform dimensionless processing on the depth gradient distribution data to generate a standardized geometric descriptor; concatenate the standardized texture descriptor and the standardized geometric descriptor according to a preset dimensional order to form a multi-dimensional defect feature vector for the current candidate defect region.

[0064] The edge computing device performs L2 normalization on the texture feature vector fk within the region, making the vector magnitude 1, using the formula f'k = fk / |fk|. It then sums and normalizes the B frequency values ​​of the texture feature distribution histogram Hk, making their sum 1. The texture difference feature parameter dtex,k is normalized by dividing it by its statistical maximum value. The normalized data are then concatenated in the order of fk dimension, the B dimensions of Hk, and the 1 dimension of dtex,k, forming a standardized texture descriptor Φtex,k with dimensions Df+B+1. For each spatial point (Δα, Gρ) in the depth gradient distribution data Gk, Δα is normalized by dividing it by the maximum possible offset angle π, and Gρ is normalized by dividing it by the statistical maximum value of Gρ. The normalized data is then expanded into a one-dimensional vector in spatial point order, forming a standardized geometric descriptor Φgeo,k. Φtex,k and Φgeo,k are concatenated to form a multi-dimensional defect feature vector Φk=[Φtex,k,Φgeo,k].

[0065] Step S147: Call the preset defect category classifier to classify the multi-dimensional defect feature vector and generate the defect category determination result of the current candidate defect region. The defect category determination result indicates that the current candidate defect region belongs to the corresponding defect type in the preset defect type set.

[0066] The defect category classifier adopts a fully connected neural network structure, consisting of two fully connected layers and one softmax output layer. The weight matrix of the first fully connected layer is Wclass1, the bias vector is bclass1, and the output is hclass1=ReLU(Wclass1×Φk+bclass1), with a length of Cclass1. The weight matrix of the second fully connected layer is Wclass2, the bias vector is bclass2, and the output is hclass2=Wclass2×hclass1+bclass2, with a length of hclass2 equal to the preset number of defect types Ctype. hclass2 is input into the softmax function, which outputs the predicted probability of each category, Ptype(c)=exp(hclass2[c]) / ∑(exp(hclass2[j])), summed and iterated through j=1 to Ctype. The category index corresponding to the maximum value of Ptype(c) is taken as the defect category determination result ck for the current candidate defect region.

[0067] Step S148: Based on the spatial distribution of abnormal depth abrupt change regions in the depth gradient distribution data, generate the defect geometric morphology parameters of the current candidate defect region. The defect geometric morphology parameters include the length extension direction, width expansion direction, and depth indentation degree of the defect region.

[0068] Starting from the spatial distribution of the abnormal depth mutation region extracted in step S143, the edge computing device calculates the azimuth and elevation angles of the major axis direction vector vlen of the minimum circumscribed ellipsoid of the region as the length extension direction, the azimuth and elevation angles of the minor axis direction vector vwid as the width extension direction, and the magnitude of the normal depth vector vdep as the depth indentation degree. The three together constitute the defect geometric morphology parameters.

[0069] Step S150: Based on the defect category determination results and defect geometric morphology parameters, generate a comprehensive defect detection report for the battery module to be inspected, and output the comprehensive defect detection report to the production management terminal to trigger the defective product sorting operation.

[0070] Step S151: Collect the defect category determination results and defect geometric morphology parameters of all candidate defect regions in the 3D candidate defect region distribution map, and construct a summary set of defect information for the battery module to be tested.

[0071] The edge computing device traverses K candidate defect regions and summarizes the index k, defect category determination result ck, defect geometric morphology parameter Gparam,k, boundary contour coordinate information and region volume parameter of each candidate defect region into a defect record. All K defect records constitute the defect information summary set S={(k,ck,Gparam,k,Bboxk,Vk)}, where Bboxk is the boundary contour coordinate information of the k-th candidate defect region and Vk is the region volume parameter of the k-th candidate defect region.

[0072] Step S152: Based on the defect category determination results of each candidate defect area in the defect information summary set, statistically analyze the distribution of the number of defect types and the frequency distribution of each defect type in the battery module to be tested.

[0073] The edge computing device iterates through all records in the defect information aggregation set S, counts the number of times each defect category c appears (Nc), and forms a defect category frequency distribution {Ctype, c, Nc}, where Ctype and c are the category names, and Nc is the total number of times that category appears. The number of defect types is the total number of categories with Nc greater than 0.

[0074] Step S153: Based on the defect geometric morphology parameters of each candidate defect region in the defect information summary set, perform geometric morphology parameter distribution analysis on the candidate defect regions with the same defect type to generate typical geometric morphology features and geometric parameter variation ranges for each defect type.

[0075] For each defect type c, candidate defect regions with defect category determination results of c are selected from S, and the depth indentation component of their defect geometry parameters is extracted. The mean μdep,c and standard deviation σdep,c of this component are calculated. The mean is taken as the typical depth indentation degree of this defect type, and the interval [μdep,c-2σdep,c,μdep,c+2σdep,c] is taken as the typical variation range of the depth parameter. For the length extension direction and the width expansion direction, the spherical mean direction and the direction concentration of the direction vector are calculated respectively. The spherical mean direction is taken as the typical direction feature, and the direction concentration is taken as the variation range of the direction parameter.

[0076] Step S154: Perform spatial distribution analysis on the defect geometric morphology parameters of all candidate defect regions in the defect information summary set to determine the spatial clustering area and spatial dispersion degree of different defect types on the surface of the battery module to be tested.

[0077] The edge computing device groups the center coordinates of all candidate defect regions in the defect information aggregation set S according to defect type. For each defect type c, a density-based noise-based spatial clustering algorithm is used to cluster the set of its center coordinates, with the cluster radius set to rcluster. Each cluster obtained after clustering corresponds to a spatial aggregation region, and the coverage area of ​​the cluster is the boundary of the spatial aggregation region. The average Euclidean distance from each point within a cluster to the centroid of the cluster is calculated as the degree of spatial dispersion.

[0078] Step S155: Generate defect distribution statistical description data based on the category quantity distribution of defect types, the frequency distribution of each defect type, the typical geometric morphological characteristics of each defect type, and the spatial aggregation areas of different defect types on the surface of the battery module to be inspected.

[0079] The edge computing device integrates the defect category frequency distribution in step S152, the typical geometric morphological features and geometric parameter variation range in step S153, and the spatial clustering area and spatial dispersion degree in step S154 into a structured defect distribution statistical description data.

[0080] Step S156: Based on the defect distribution statistical description data, match the corresponding defect grade scoring threshold from the preset defect grade evaluation standard library to generate the defect grade scoring result of the battery module to be tested.

[0081] The preset defect level assessment standard library is a rule matching table. The condition part of each rule includes the frequency of defect type occurrence, the proportion of the total volume of the defect area to the surface area of ​​the battery module, and the spatial dispersion of the defect. The conclusion part is the quality rating. The edge computing device matches the defect distribution statistical description data obtained in step S155 with the condition part of each rule one by one. If a match is successful, the corresponding quality rating is output as the defect quality rating result.

[0082] Step S157: Combine the defect information summary set, defect distribution statistical description data, and defect grade score results to form a comprehensive defect detection report.

[0083] The edge computing device serializes the above three parts of data into a structured data format, forming a comprehensive defect detection report. This report is sent to the production management terminal, which analyzes the defect grade score in the report. If the grade score is lower than the acceptable product threshold, a rejection instruction is sent to the sorting robot arm.

[0084] Step S210: Obtain intermediate feature data generated by the defect region pre-segmentation network during the candidate region generation process. The intermediate feature data includes the deep texture fusion feature volume and the generation parameters of multiple 3D candidate regions.

[0085] During step S133, the edge computing device saves the deep texture fusion feature volume Ffused output by the intermediate layer of the defect region pre-segmentation network and the generation parameters of all M three-dimensional candidate regions output by the region candidate generation module as intermediate feature data.

[0086] Step S211: Perform feature activation region localization processing on the deep texture fusion feature volume to generate a set of spatial coordinates of the high activation response regions in the deep texture fusion feature volume.

[0087] The edge computing device takes the maximum response value at each spatial location along the channel dimension of Ffused and generates a response intensity map Rspatial. Spatial locations in Rspatial where the response value exceeds the activation threshold τactivate are marked as high-activation response regions, and their spatial coordinates are recorded to form a spatial coordinate set Pactivate.

[0088] Step S212: Based on the set of spatial coordinates, extract the local three-dimensional surface representation image patch corresponding to the high activation response region from the three-dimensional surface representation image.

[0089] Centered on each spatial coordinate in Pactivate, a local 3D surface representation image patch of size Lpatch×Lpatch×Lpatch is extracted from the 3D surface representation image V, where Lpatch is a preset patch size. The set of all extracted image patches is denoted as {Vpatch, i}.

[0090] Step S213: Perform defect detail enhancement processing on the local three-dimensional surface characterization image block to generate a high-resolution defect detail image block.

[0091] The edge computing device upsamples each local 3D surface representation image patch Vpatch,i through a pre-trained super-resolution reconstruction network. The super-resolution reconstruction network adopts a 3D subpixel convolutional neural network structure to increase the spatial resolution of the input image patch by r times, generating a high-resolution defect detail image patch Vhigh,i.

[0092] Step S214: Use high-resolution defect detail image patches as new training samples to update the weight parameters of the defect region pre-segmentation network.

[0093] The edge computing device uses {Vhigh, i} as a new training sample, merges it with the original training sample set, and performs fine-tuning training on the defect region pre-segmentation network. During fine-tuning, a smaller learning rate is used, along with the same loss function and optimizer as the original training. After several iterations, the updated weight parameters of the defect region pre-segmentation network are obtained.

[0094] Step S215: Use the updated defect region pre-segmentation network to perform potential defect region localization processing on the three-dimensional surface characterization image of the battery module to be inspected.

[0095] After the battery module to be inspected arrives at the inspection station, its three-dimensional surface characterization image is sent to the updated defect region pre-segmentation network, and the processing steps S131 to S137 are performed to generate a more accurate three-dimensional candidate defect region distribution map.

[0096] Step S216: Based on the erroneous classification results in the defect category determination results, backpropagate the classification error to adjust the decision boundary parameters of the defect category classifier.

[0097] When the results of manual re-inspection reported by the production management terminal are inconsistent with the defect category determination results, the inconsistent samples are recorded as misclassification results. The edge computing device uses the cross-entropy loss between the correct category label of the misclassified result and the predicted probability of the defect category classifier as a supervision signal, and updates the fully connected layer weight parameters Wclass1, Wclass2 and bias parameters bclass1, bclass2 of the defect category classifier in step S147 through the error backpropagation algorithm, so that the decision boundary is adjusted towards the correct classification direction.

[0098] Step S310: Obtain the process parameter record data of the production process of the battery module to be tested within a preset historical time period. The process parameter record data includes the welding temperature sequence, welding pressure sequence, and liquid injection flow rate sequence.

[0099] Edge computing devices establish a data interface connection with the manufacturing execution system database in the production workshop. Through this interface, they can query the process parameter records of the battery module under test in the upstream welding and electrolyte injection processes. The welding temperature sequence is {Ttemp(t)}, where t is the index of the time sampling point when the battery module passes through the welding station; the welding pressure sequence is {Fpress(t)}; and the electrolyte injection flow rate sequence is {Vflow(t)}.

[0100] Step S311: Perform time-series correlation analysis on the defect category determination results in the defect detection comprehensive report and the process parameter record data to generate a correlation matrix between defect category and process parameter fluctuation.

[0101] The edge computing device matches the defect category determination results with the corresponding process positions on the battery module according to the spatial location of the defect, and extracts the time series segment of the process parameters at that position. The Pearson correlation coefficient between the occurrence probability of each defect type c and the fluctuation of each process parameter sequence is calculated. The fluctuation is quantified by the standard deviation of the sequence. The correlation coefficients constitute the correlation degree matrix Rcorr(c, p), where p is the process parameter index.

[0102] Step S312: Based on the correlation matrix, identify the associated process parameter items and process parameter fluctuation thresholds that lead to specific defect types.

[0103] For elements in the correlation matrix whose absolute correlation coefficient exceeds the correlation threshold τcorr, the corresponding process parameter item is marked as associated with a specific defect type. Statistical analysis is performed on the fluctuation value of this associated process parameter item, and the upper quantile of the parameter fluctuation in the normal product sample is taken as the process parameter fluctuation threshold τparam.

[0104] Step S313: When the real-time monitoring value of the associated process parameter exceeds the process parameter fluctuation threshold, a process parameter adjustment instruction is generated and fed back to the control unit of the corresponding production process.

[0105] Edge computing devices continuously receive real-time process parameter monitoring values ​​pushed by the manufacturing execution system. When the real-time fluctuation of a certain related process parameter exceeds its corresponding process parameter fluctuation threshold τparam, a process parameter adjustment instruction is generated. The instruction includes the target parameter identifier and the target adjustment value, and is sent to the programmable logic controller of the corresponding process through the data interface.

[0106] For example, step S410: Construct a three-dimensional standard model of the battery module to be tested. The three-dimensional standard model contains the ideal depth dimension information and ideal texture dimension information of the standard battery module.

[0107] The edge computing device obtains the computer-aided design 3D model data of the standard battery module from the product design database, converts the computer-aided design model into the same spherical mapping space representation as in step S120, extracts the ideal normal direction vector nideal and the ideal radius of curvature ρideal of the standard battery module surface as ideal depth dimension information, extracts the ideal texture primitive code LBPideal and the ideal texture continuity coefficient cideal of the standard battery module surface material as ideal texture dimension information, and constructs the 3D standard model Vstd.

[0108] Step S411: Perform spatial registration processing on the three-dimensional surface representation image and the three-dimensional standard model to align the three-dimensional surface representation image and the three-dimensional standard model to a unified spatial coordinate system.

[0109] Edge computing devices employ an iterative nearest-point algorithm to rigidly register the point cloud of a 3D surface representation image V with the point cloud of a 3D standard model Vstd. The registration process involves iteratively solving for the rotation matrix Rreg and translation vector Treg by minimizing the sum of squared Euclidean distances between corresponding points in the two point clouds. The coordinates of each spatial point in V are then transformed to the coordinate system of Vstd using Rreg and Treg, completing the spatial registration.

[0110] Step S412: Calculate the point-by-point depth deviation between the depth dimension information of the three-dimensional surface representation image and the ideal depth dimension information, and normalize the depth deviation to generate a standardized depth deviation distribution map.

[0111] For each corresponding position in the registered space, calculate the normal direction deviation angle Δαbias = arccos(n·nideal) and the radius of curvature deviation Δρbias = |ρ - ρideal|. Normalize Δαbias by dividing by π and Δρbias by dividing by ρideal, and take the weighted sum of the two as the depth deviation for that point. Arrange the depth deviations of all spatial points according to their spatial positions to form a standardized depth deviation distribution map.

[0112] Step S413: Calculate the texture primitive encoding deviation between the texture dimension information of the three-dimensional surface representation image and the ideal texture dimension information, and normalize the texture primitive encoding deviation to generate a standardized texture deviation distribution map.

[0113] For each corresponding position in the registered space, the Hamming distance of the texture primitive encoding is calculated as the texture primitive encoding bias ΔLBP. ΔLBP is then normalized by dividing by the maximum possible Hamming distance P, where P is the number of binary bits in the texture primitive encoding. The normalized texture primitive encoding biases of all spatial points are arranged according to their spatial location to form a standardized texture bias distribution map.

[0114] Step S414: The standardized depth deviation distribution map and the standardized texture deviation distribution map are weighted and fused according to the preset weight coefficients to generate an overall deviation distribution map between the battery module to be tested and the three-dimensional standard model.

[0115] The edge computing device takes the deviation value of each point in the standardized depth deviation distribution map, multiplies it by the weighting coefficient wdeep, and takes the deviation value of each point in the standardized texture deviation distribution map, multiplies it by the weighting coefficient wtex, and adds wdeep + wtex ​​= 1. The overall deviation value is obtained by summing the values ​​of each point. The overall deviation value of all spatial points constitutes the overall deviation distribution map.

[0116] Step S415: Based on the overall deviation distribution map, perform candidate region expansion processing on the three-dimensional candidate defect region distribution map generated by the defect region pre-segmentation network, and supplement the regions whose overall deviation exceeds the preset deviation threshold as new candidate defect regions.

[0117] The edge computing device extracts continuous spatial regions in the overall deviation distribution map where the deviation value exceeds the deviation threshold τbias, calculates the anchor point coordinates, region size parameters, and internal texture feature vectors of the circumscribed cuboid region, and adds them as new candidate defect regions to the 3D candidate defect region distribution map.

[0118] Step S510: Collect comprehensive defect detection reports for multiple battery modules to be tested within a preset time period and construct a historical database of defect detection results.

[0119] The edge computing device persistently stores the defect information summary set S of each battery module processed within a preset time period, the defect distribution statistical description data, and the defect grade rating results, and constructs a historical database of defect detection results indexed by the unique identifier of the battery module and the timestamp.

[0120] Step S511: Perform cluster analysis on the defect geometric morphology parameters in the historical database of defect detection results to generate several typical defect morphology clusters, each of which corresponds to a cluster center morphology parameter.

[0121] The edge computing device extracts the defect geometric morphology parameters of all candidate defect regions from the historical database of defect detection results. It then concatenates the depth of indentation, length extension direction vector, and width expansion direction vector of each candidate defect region into a morphological feature vector. The K-means clustering algorithm is used to cluster all morphological feature vectors, with a preset number of clusters, Kcluster. After clustering converges, Kcluster of cluster center morphological parameters are obtained.

[0122] Step S512: Extract the defect category determination results of each battery module to be tested from the historical database of defect detection results, and establish the prior probability distribution of defect categories associated with each typical defect morphology cluster.

[0123] For the m-th typical defect morphology cluster, the frequency of occurrence of each defect category in all candidate defect regions belonging to this cluster is counted, and the prior probability distribution of the defect category Pprior(c|m) is obtained after normalization.

[0124] Step S513: Use the cluster center morphological parameters of typical defect morphological clusters and the prior probability distribution of defect categories as prior knowledge parameters.

[0125] Edge computing devices store the morphological parameters of the Kcluster cluster centers and the corresponding prior probability distributions of the Kcluster defect categories as a prior knowledge parameter library.

[0126] Step S514: When performing defect detection on a new battery module to be tested, the prior knowledge parameters are compared and analyzed with the defect geometric morphology parameters of the current candidate defect region in the three-dimensional candidate defect region distribution map to correct the defect category judgment result output by the defect category classifier and generate the corrected defect category judgment result.

[0127] The edge computing device calculates the Euclidean distance between the geometric morphological parameters of the current candidate defect region and the morphological parameters of each of the Kcluster cluster centers, selects the index mnearest of the nearest typical defect morphological cluster, and retrieves the corresponding prior probability distribution Pprior(c|mnearest) of the defect category. The predicted probability Ptype(c) output by the defect category classifier in step S147 is multiplied element-wise with Pprior(c|mnearest), and then normalized to obtain the corrected predicted probability Pcorrected(c) = (Ptype(c) × Pprior(c|mnearest)) / ∑(Ptype(j) × Pprior(j|mnearest)), and the summation iterates through j=1 to Ctype. The category corresponding to the maximum value of Pcorrected(c) is taken as the corrected defect category determination result.

[0128] Step S610: Obtain the initial light intensity distribution parameters of the illumination source of the battery module to be tested in the multi-angle imaging device.

[0129] The edge computing device reads the initial drive current value of each light source unit from the drive controller register of the lighting source, converts the drive current value into a light intensity value according to the current and light intensity calibration curve of the light source unit, and arranges the light intensity values ​​of all light source units according to their spatial positions to form the initial light intensity distribution parameter L0.

[0130] Step S611: Perform brightness uniformity analysis on each surface image unit of the battery module in the appearance image set to generate a local brightness deviation distribution map of each surface image unit of the battery module.

[0131] For the nth battery module surface image unit In, the edge computing device divides the image into a uniform grid, with each grid unit being the size of gridsize multiplied by gridsize pixels. The mean grayscale value of all pixels within each grid unit is calculated. The arithmetic mean of the overall grid mean is used as the overall brightness reference value Iref. The local brightness deviation of each grid unit is the grid mean minus Iref. The local brightness deviations of all grid units are arranged according to their grid positions to form a local brightness deviation distribution map Bn.

[0132] Step S612: Based on the local brightness deviation distribution map, generate a light source brightness compensation command for the lighting source. The light source brightness compensation command is used to instruct the light source units at different positions to adjust the output light intensity to balance the brightness distribution of the image units on the surface of the battery module.

[0133] The edge computing device maps Bn back to the array of light source units of the illumination source through the optical path mapping relationship between the light source and the image. The optical path mapping relationship is pre-determined during the system calibration phase by sequentially illuminating individual light source units and acquiring the response distribution of the corresponding images. For each light source unit, the average local brightness deviation of the grid within its mapping area is taken. Negative deviations correspond to increased output light intensity for the light source unit, and positive deviations correspond to decreased output light intensity for the light source unit. The adjustment amount is determined by multiplying the absolute value of the deviation by a preset gain coefficient, gainlight. The generated adjustment command includes the identifier of each light source unit and the adjusted target current value.

[0134] Step S613: Send the light source brightness compensation command to the driver controller of the lighting source to trigger the light source parameter adjustment operation.

[0135] The edge computing device sends the light source brightness compensation command to the driver controller of the lighting source through a serial communication interface. After parsing the command, the driver controller adjusts the output current of the corresponding light source unit to complete the adjustment of the light source parameters.

[0136] Step S614: Obtain the set of secondary imaging images after adjusting the light source parameters, perform image quality improvement and fusion processing on the set of secondary imaging images and the set of appearance images, and generate an optimized set of battery module surface image units.

[0137] After the light source parameters are adjusted, the edge computing device triggers the multi-angle imaging device to synchronously acquire images, resulting in a set of secondary imaging images {I′1, I′2, ..., I′N}. For each pair of original images In and secondary imaging images I′n, the local signal-to-noise ratio (SNR) is used as the fusion weight. The SNR is defined as the mean of the local region divided by the standard deviation. For each pixel position in the image, the ratio of the SNR within the local window centered on that pixel in the two images is taken as the fusion weight wSNR. The fused pixel value is (1-wSNR)×In(i,j)+wSNR×I′n(i,j). All fused images constitute the optimized set of battery module surface image units {Iopt,n}.

[0138] Step S615: Update the appearance image set using the optimized battery module surface image unit set, and re-execute the view association mapping process to generate an optimized three-dimensional surface characterization image.

[0139] The edge computing device replaces the original {I1, I2, ..., IN} with {Iopt, n} and re-executes all sub-steps of step S120 to generate an optimized 3D surface representation image Vopt. Subsequent processing uses Vopt instead of V.

[0140] Step S710: Perform texture direction field estimation processing on the texture dimension information in the three-dimensional surface characterization image, calculate the principal direction angle of the texture primitive encoding of each spatial point in the texture dimension information in the local neighborhood, and generate the texture principal direction field distribution map of the surface of the battery module to be detected.

[0141] For each spatial point in the texture dimension information T, the edge computing device takes the gradient of the texture primitive encoding in its local neighborhood. The gradient calculation is approximated by the Sobel operator on a discrete grid in the spherical mapping space. The gradient direction is taken as the arctangent value arctan(Gy / Gx) as the principal direction angle θtex of the texture at that spatial point. The θtex of all spatial points constitutes the principal direction field distribution map Ψ of the texture.

[0142] Step S711: Based on the texture principal direction field distribution map, identify the location of the texture direction break line where the texture principal direction on the surface of the battery module to be detected changes abruptly. The location of the texture direction break line is the position where the texture principal direction angle deviation between adjacent spatial points exceeds the preset angle change threshold.

[0143] The edge computing device calculates the texture principal direction angle deviation Δθ = |θ1 - θ2| for each pair of adjacent spatial points in Ψ. If Δθ exceeds the angle abrupt change threshold θbreak, the midpoint coordinates of the two spatial points are recorded as the texture direction breakpoint. The spatially adjacent texture direction breakpoints are connected into continuous line segments through curve fitting to obtain the texture direction breakpoint location set {Lbreak}.

[0144] Step S712: Extract the spatial coordinate sequence of the texture direction break line position, and perform spatial overlap calculation on the spatial coordinate sequence of the texture direction break line position and the boundary contour coordinate information of the candidate defect area in the three-dimensional candidate defect area distribution map to generate a spatial consistency mapping relationship between the texture break line and the boundary of the candidate defect area.

[0145] For each fracture line along a texture direction, the ratio of the intersection length of its spatial coordinate sequence with the boundary contour of each candidate defect region to the total length of the fracture line is calculated as the spatial overlap. Fracture lines with a spatial overlap exceeding the overlap threshold τoverlap are paired with candidate defect regions to establish a spatial consistency mapping relationship.

[0146] Step S713: Based on the spatial consistency mapping relationship, perform defect type tendency weighting processing on candidate defect regions whose spatial overlap with the fracture line in the texture direction exceeds the preset overlap threshold, and increase the tendency weight of the candidate defect region to be judged as a crack-type defect.

[0147] Before the softmax layer of the defect category classifier in step S147, the edge computing device adds a tendency increment Δwcrack to the score value corresponding to the crack defect type. Δwcrack is proportional to the spatial overlap. After increasing it, the softmax output is recalculated, and the prediction probability of the crack category is increased accordingly.

[0148] Step S714: Perform directional gradient magnitude calculation on the texture main direction field distribution map to generate a texture direction change intensity distribution map. The value of each spatial point in the texture direction change intensity distribution map reflects the intensity of texture direction change in the neighborhood of that spatial point.

[0149] For each spatial point in Ψ, the edge computing device calculates the standard deviation σlocal of the texture direction angle within its neighborhood, using σlocal as the value of the drastic change in direction at that spatial point. The σlocal values ​​of all spatial points constitute a distribution map Ω of the drastic change in texture direction.

[0150] Step S715: Normalize the distribution map of the intensity of texture direction changes to make its numerical scale match the depth dimension information; use the normalized distribution map of the intensity of texture direction changes as an additional input dimension, and perform dimension overlay processing with the depth dimension information and texture dimension information to generate a three-dimensional surface representation image with extended dimensions.

[0151] The edge computing device performs Min-Max normalization on Ω to make its numerical range consistent with the normalized radius of curvature of the depth data layer Dlayer. The normalized Ω is then appended as a new data layer to the 3D surface representation image V, generating an extended-dimensional 3D surface representation image Vext, whose data layers include a depth data layer, a texture data layer, and a texture orientation variation layer.

[0152] Step S716: Input the extended-dimensional 3D surface representation image into the defect region pre-segmentation network to replace the original 3D surface representation image for potential defect region localization processing, and generate a 3D candidate defect region distribution map that supplements the sensitivity of texture direction anomalies.

[0153] The edge computing device feeds Vext into the multi-scale feature encoding module of the defect region pre-segmentation network. A third parallel branch is added to the existing depthwise convolutional and texture convolutional branches to handle texture orientation variation layers. The structure of this new branch is the same as the depthwise convolutional branch, outputting multi-scale texture orientation feature maps {Fo1, Fo2, Fo3}. In the feature interaction fusion module, cross-attention fusion of texture orientation features with depth features and texture features is added to generate an expanded fused feature volume. Subsequent steps S133 to S137 remain unchanged, and the generated 3D candidate defect region distribution map includes texture orientation-sensitive areas.

[0154] Step S810: Extract the boundary contour coordinate information of each candidate defect region in the 3D candidate defect region distribution map, perform 3D boundary contour parametric modeling on the boundary contour coordinate information, and generate a 3D boundary contour parametric model for each candidate defect region. The 3D boundary contour parametric model uses the boundary contour curvature function and the boundary contour torsion function as description parameters.

[0155] The edge computing device samples the boundary contour of each candidate defect region according to the arc length parameter s, obtaining a boundary point sequence P(s). For each sampled point, the curvature κ(s) and torsion τ(s) of the boundary contour at that point are calculated: curvature κ(s) = |P'(s)×P””s)| / |P'(s)|^3, and torsion τ(s) = (P'(s)×P””s)·P”’(s) / |P'(s)×P””s)|^2, where P'(s), P””s, and P”’(s) represent the first, second, and third derivatives of P(s) with respect to the arc length s, respectively, and are discretized using the finite difference method. (κ(s), τ(s)) at all arc length positions constitute a three-dimensional boundary contour parameterized model.

[0156] Step S811: Based on the three-dimensional boundary profile parameterization model, calculate the curvature change rate sequence of the boundary profile curvature function along the boundary arc length and the deflection change rate sequence of the boundary profile torsion function along the boundary arc length for each candidate defect region.

[0157] The edge computing device takes the first-order difference of the curvature function κ(s) along the arc length s to obtain the curvature change rate sequence {dκ / ds}. Similarly, it takes the first-order difference of the torsion function τ(s) along the arc length s to obtain the torsion change rate sequence {dτ / ds}.

[0158] Step S812: Input the curvature change rate sequence into the preset time-series autoencoder for encoding processing to generate a time-series feature encoding vector of boundary curvature change; input the torsion change rate sequence into the time-series autoencoder for encoding processing to generate a time-series feature encoding vector of boundary torsion change.

[0159] The temporal pattern autoencoder uses a long short-term memory (LSTM) network as the basic unit for both the encoder and decoder. The encoder feeds the input sequence into the LTM unit step by step, and the hidden state vector hcurv of the last time step serves as the temporal feature encoding vector for the boundary curvature change. Similarly, the torsion rate change sequence is fed into another LTM encoder with the same structure to obtain the temporal feature encoding vector htors for the boundary torsion change.

[0160] Step S813: Standardize the temporal feature encoding vectors of boundary curvature change and boundary torsion change respectively to make them have a unified numerical scale; concatenate and fuse the standardized vectors to generate a boundary geometric invariance feature vector describing the three-dimensional boundary morphology of the candidate defect region.

[0161] The edge computing device performs Z-score normalization on hcurv and htors respectively. After normalization, the mean of the vector components is 0 and the standard deviation is 1. The two normalized vectors are concatenated end to end to obtain the boundary geometric invariance eigenvector hgeom=[hcurv, htors].

[0162] Step S814: Perform principal component analysis to reduce the dimensionality of the boundary geometric invariant eigenvectors, generating low-dimensional boundary geometric invariant eigenvectors.

[0163] The edge computing device uses the transformation matrix Wpca, which is calculated in advance on a large number of samples, to perform a linear transformation on hgeom. The first Dpca principal components are taken to form a low-dimensional boundary geometric invariant eigenvector hgeom_reduced=Wpca×hgeom, where Dpca is the preset target dimension.

[0164] Step S815: Call the preset defect boundary geometry pattern classifier to classify the low-dimensional boundary geometry invariant feature vector and generate the boundary geometry pattern classification result. The boundary geometry pattern classification result indicates that the boundary geometry of the current candidate defect region belongs to the corresponding category in the preset boundary geometry pattern category.

[0165] The defect boundary geometric pattern classifier employs a support vector machine (SVM) model. During the training phase, the hyperplane parameters are trained using samples labeled with boundary geometric pattern category tags. The edge computing device inputs `hgeom_reduced` into this SVM, and after kernel function mapping, calculates the distance to the hyperplane of each pattern category. The index of the pattern category with the largest distance is output as the boundary geometric pattern classification result `mgeom`.

[0166] Step S816: Based on the boundary geometry pattern classification result, perform boundary geometry constraint correction processing on the defect category determination result. When there is a category conflict between the boundary geometry pattern classification result and the defect category determination result, the boundary geometry pattern classification result shall be used as the final defect category determination result.

[0167] The edge computing device reads a preset mapping table between boundary geometry patterns and defect categories, and looks up the corresponding expected defect category cgeom based on mgeom. If the defect category determination result ck output by the defect category classifier in step S147 is inconsistent with cgeom, a category conflict is determined, and the final defect category determination result is corrected to cgeom.

[0168] Step S910: Perform layered slicing processing on the depth dimension information in the three-dimensional surface characterization image. Divide the depth dimension information into several depth layers along the normal direction of the surface of the battery module to be detected. Each depth layer contains a subset of depth data at the spatial location corresponding to that depth layer.

[0169] The edge computing device determines the main direction of slicing based on the average normal direction of the overall point cloud on the surface of the battery module to be detected. Along this main direction, the depth dimension information D is divided into L depth layers according to the layer spacing Δz. The l-th depth layer contains the depth data corresponding to all spatial points whose depth coordinates fall within the interval [zmin+(l-1)×Δz, ​​zmin+l×Δz], where l=1, 2, ..., L.

[0170] Step S911: Perform depth anomaly region detection processing on the depth data subset within each depth layer, identify depth anomaly patch regions within each depth layer whose depth values ​​deviate from the depth mean of that depth layer, and generate a depth anomaly patch distribution map corresponding to each depth layer.

[0171] For the l-th depth layer, calculate the mean μρ,l and standard deviation σρ,l of the radius of curvature of all spatial points within the layer. Spatial points whose radius of curvature deviates from the mean by more than cslice times the standard deviation are marked as depth anomalies, where cslice is a preset multiple. Perform connectivity component analysis on the depth anomalies, aggregating spatially adjacent depth anomalies into depth anomaly patch regions. The set of bounding rectangles of all depth anomaly patch regions in this layer constitutes the depth anomaly patch distribution map SliceAnom,l for this depth layer.

[0172] Step S912: Extract the distribution map of depth anomalies between adjacent depth layers, perform cross-layer connectivity analysis on the distribution maps of depth anomalies between adjacent depth layers, identify the sequence of depth anomalies that have spatial connectivity in continuous depth layers, and generate a set of cross-layer depth anomaly patch connectivity.

[0173] The edge computing device starts at l=1 and sequentially checks the spatial overlap between the patch regions in the l-th layer depth anomaly patch distribution map SliceAnom,l and the (l+1)-th layer SliceAnom,l+1. If the spatial intersection-union ratio of patch A in the l-th layer and patch B in the (l+1)-th layer is greater than the connectivity threshold τconn, then A and B are determined to have a cross-layer connectivity relationship. All patch sequences with connectivity relationships are traced along the depth direction to form a cross-layer depth anomaly patch connectivity set {Clusterj}, j=1,2,…,J.

[0174] Step S913: Perform three-dimensional morphological feature extraction processing on each cross-layer depth anomalous patch connected body in the cross-layer depth anomalous patch connected body set to generate three-dimensional volume morphological features of each cross-layer depth anomalous patch connected body. The three-dimensional volume morphological features include the longitudinal extension depth, lateral expansion area and three-dimensional shape factor of the connected body.

[0175] For each cross-layer depth anomaly patch connected component Clusterj, calculate the span of its contained spatial points in the depth direction as the longitudinal extension depth Dextend,j. Calculate the pixel area it occupies on the projection plane perpendicular to the depth direction as the lateral extension area Area,j. Calculate its three-dimensional shape factor SFj = 36 × π × Vj^2 / Sj^3, where Vj is the volume of the connected component and Sj is the surface area of ​​the connected component. The three-dimensional volumetric morphological feature is composed of the above three parameters Φ3D,j = (Dextend,j, Area,j, SFj).

[0176] Step S914: Perform dimensionless normalization on the longitudinal extension depth, lateral expansion area and three-dimensional shape factor in the three-dimensional volume morphology features to generate a standardized geometric morphology feature vector; perform normalization on the texture feature vector inside the candidate defect region; and concatenate the standardized geometric morphology feature vector with the standardized texture feature vector to generate a multi-level defect representation vector that integrates depth layering information.

[0177] The edge computing device normalizes each component of Φ3D,j by dividing it by the statistical maximum value of the corresponding component to obtain a standardized geometric feature vector Φ'3D,j. The texture feature vector fk within the candidate defect region is L2 normalized to obtain f'k. Φ'3D,j and f'k are concatenated to form a multi-level defect representation vector Ψj=[Φ'3D,j,f'k].

[0178] Step S915: Call the preset multi-level defect classifier to classify the multi-level defect representation vector and generate a multi-level defect category determination result that integrates deep hierarchical discrimination information.

[0179] The multi-level defect classifier employs a fully connected neural network with the same structure as the defect category classifier in step S147, but uses training samples containing depth layering information during training. The edge computing device inputs Ψj into this classifier, which outputs the predicted probabilities for each category after passing through two fully connected layers and a softmax layer. The category index corresponding to the maximum probability is taken as the multi-level defect category determination result cmulti,j. This result integrates the spatial distribution features of the defect region across multiple depth layers, achieving higher discrimination accuracy for defects with significant depth extension features.

[0180] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0181] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, it is intended that all variations falling within the meaning and scope of equivalents of the claims be included within the present invention.

Claims

1. A battery module defect detection method based on deep learning, characterized in that, The method includes: A set of appearance images of the battery module to be tested is obtained by a preset multi-angle imaging device. The set of appearance images includes multiple battery module surface image units that are synchronously acquired from different imaging perspectives. Each battery module surface image unit has a view orientation mark and acquisition time mark corresponding to the imaging perspective. Viewpoint association mapping processing is performed on the set of appearance images. The surface image units of the battery module with different viewpoint orientation marks are projected onto a unified spherical mapping space. The viewpoint association correspondence between the surface image units of each battery module is established, and a three-dimensional surface characterization image of the battery module to be detected in the spherical mapping space is generated. The three-dimensional surface characterization image contains depth dimension information reflecting the curved surface structure of the battery module and texture dimension information reflecting the surface material properties of the battery module. A preset defect region pre-segmentation network is invoked to perform potential defect region localization processing on the three-dimensional surface characterization image, generating a three-dimensional candidate defect region distribution map containing several candidate defect regions. Each candidate defect region in the three-dimensional candidate defect region distribution map has boundary contour coordinate information, region volume parameters, and region internal texture feature vector. Based on the depth dimension information and the texture dimension information, multi-dimensional defect feature analysis processing is performed on each candidate defect region in the three-dimensional candidate defect region distribution map to generate defect category determination results and defect geometric morphology parameters for each candidate defect region. The defect category determination results are generated by calling a preset defect category classifier. Based on the defect category determination result and the defect geometric morphology parameters, a comprehensive defect detection report for the battery module to be inspected is generated, and the comprehensive defect detection report is output to the production management terminal to trigger the defective product sorting operation.

2. The battery module defect detection method based on deep learning according to claim 1, characterized in that, The step of performing viewpoint association mapping processing on the set of appearance images involves projecting the surface image units of the battery module with different viewpoint orientation markers onto a unified spherical mapping space, establishing viewpoint association correspondences between the surface image units of each battery module, and generating a three-dimensional surface representation image of the battery module to be detected in the spherical mapping space, including: Read the view orientation mark corresponding to each surface image unit of the battery module one by one from the set of appearance images, and parse the optical center coordinates and imaging optical axis direction vector of the imaging camera corresponding to the view orientation mark in the preset spatial coordinate system. Based on the optical center coordinates and the imaging optical axis direction vector, a ray inverse tracing mapping relationship is constructed between the pixels of each battery module surface image unit and the spatial points on the preset spherical mapping space. The preset spatial coordinate transformation algorithm is invoked, and based on the ray inverse tracing mapping relationship, the two-dimensional pixel coordinates on the surface image unit of each battery module are transformed to the three-dimensional spherical coordinates of the preset spherical mapping space, generating a spherical projection point cloud set corresponding to each surface image unit of the battery module. A point cloud fusion process is performed on the spherical projection point cloud set generated by all battery module surface image units to eliminate coordinate overlap and density differences between projection point clouds from different viewpoints, and to generate a fused overall point cloud of the battery module surface. Extract the spatial coordinate distribution features of the overall point cloud on the surface of the battery module, perform spatial surface fitting processing on the overall point cloud on the surface of the battery module, and generate depth dimension information characterizing the surface structure of the battery module. The depth dimension information includes the normal direction information and radius of curvature information of each spatial point on the overall point cloud on the surface of the battery module. Texture primitive extraction is performed on each surface image unit of the battery module to obtain a local texture primitive encoding set for each surface image unit of the battery module. Each texture primitive encoding in the local texture primitive encoding set represents the texture mode description information of a local pixel region. Based on the viewpoint association relationship, the local texture primitive encoding set of each battery module surface image unit is mapped to the corresponding spatial point of the overall point cloud of the battery module surface to generate texture dimension information. The texture dimension information includes the texture primitive encoding and texture continuity coefficient of each spatial point on the overall point cloud of the battery module surface. The depth dimension information is standardized, the texture dimension information is normalized, and the processed depth dimension information and texture dimension information are aligned and fused in the preset spherical mapping space to generate a three-dimensional surface representation image composed of a depth data layer and a texture data layer superimposed.

3. The battery module defect detection method based on deep learning according to claim 1, characterized in that, The process of calling a preset defect region pre-segmentation network to perform potential defect region localization processing on the three-dimensional surface characterization image, generating a three-dimensional candidate defect region distribution map containing several candidate defect regions, includes: The three-dimensional surface characterization image is input into the multi-scale feature encoding module of the defect region pre-segmentation network. The multi-scale feature encoding module performs multi-scale convolution processing on the depth dimension information and texture dimension information of the three-dimensional surface characterization image to generate a depth multi-scale feature map and a texture multi-scale feature map. The deep multi-scale feature map and the texture multi-scale feature map are input into the feature interaction fusion module of the defect region pre-segmentation network. The deep multi-scale feature map and the texture multi-scale feature map are subjected to feature interaction fusion processing through a cross-attention mechanism to generate a deep texture fusion feature volume. The deep texture fusion feature volume is input into the region candidate generation module of the defect region pre-segmentation network. Multiple three-dimensional candidate regions are generated in the deep texture fusion feature volume by a three-dimensional region proposal algorithm. Each three-dimensional candidate region has anchor point coordinates, region size parameters and region boundary contour parameters. The candidate region filtering layer of the defect region pre-segmentation network performs overlap filtering and confidence ranking on the multiple three-dimensional candidate regions to filter out several candidate defect regions that meet the preset overlap conditions and preset confidence conditions. The region boundary correction process is performed on the candidate defect regions. The boundary contour of each candidate defect region is adjusted by using a three-dimensional boundary regression algorithm to generate the boundary contour coordinate information and region volume parameters of each candidate defect region. The feature pooling module of the defect region pre-segmentation network maps each candidate defect region onto the deep texture fusion feature volume, extracts the texture feature information within the corresponding spatial range of each candidate defect region, and generates the region-internal texture feature vector of each candidate defect region. Based on the boundary contour coordinates, volume parameters, and texture feature vectors of each candidate defect region, a three-dimensional candidate defect region distribution map is generated, which marks the position and contour of each candidate defect region in three-dimensional space.

4. The battery module defect detection method based on deep learning according to claim 1, characterized in that, Based on the depth dimension information and the texture dimension information, multi-dimensional defect feature analysis is performed on each candidate defect region in the three-dimensional candidate defect region distribution map to generate a defect category determination result and defect geometric morphology parameters for each candidate defect region, including: Extract the boundary contour coordinate information of the current candidate defect region from the three-dimensional candidate defect region distribution map, and extract a subset of local depth information corresponding to the current candidate defect region from the depth dimension information based on the boundary contour coordinate information; The local depth information subset is subjected to depth gradient field analysis processing to calculate the normal direction offset and gradient magnitude of each spatial point in the local depth information subset, thereby generating depth gradient distribution data that reflects the surface undulation characteristics of the current candidate defect region. Based on the depth gradient distribution data, abnormal depth abrupt change regions in the current candidate defect region whose surface undulation exceeds a preset undulation threshold are extracted, and the spatial distribution of the abnormal depth abrupt change regions is used as a component of the defect geometric morphology parameters. Extract the texture feature vector within the current candidate defect region from the three-dimensional candidate defect region distribution map, and construct the texture feature distribution histogram of the current candidate defect region based on the texture feature vector within the region. Extract the texture primitive encoding of the surrounding region adjacent to the current candidate defect region from the texture dimension information, perform texture difference comparison processing on the texture primitive encoding of the texture feature vector inside the region and the texture primitive encoding of the surrounding region, and generate texture difference feature parameters. The texture feature vector within the region, the texture feature distribution histogram, and the texture difference feature parameters are normalized to generate a standardized texture descriptor; the depth gradient distribution data is dimensionless to generate a standardized geometric descriptor; the standardized texture descriptor and the standardized geometric descriptor are concatenated according to a preset dimensional order to form a multi-dimensional defect feature vector of the current candidate defect region. A preset defect category classifier is invoked to classify the multi-dimensional defect feature vector and generate a defect category determination result for the current candidate defect region. The defect category determination result indicates that the current candidate defect region belongs to the corresponding defect type in the preset defect type set. Based on the spatial distribution of abnormal depth abrupt change regions in the depth gradient distribution data, defect geometric morphology parameters of the current candidate defect region are generated. The defect geometric morphology parameters include the length extension direction, width expansion direction, and depth indentation degree of the defect region.

5. The battery module defect detection method based on deep learning according to claim 1, characterized in that, The process of generating a comprehensive defect detection report for the battery module under test based on the defect category determination result and the defect geometric morphology parameters includes: Collect the defect category determination results and defect geometric morphology parameters of all candidate defect regions in the three-dimensional candidate defect region distribution map, and construct a summary set of defect information of the battery module to be tested. Based on the defect category determination results of each candidate defect region in the defect information summary set, the distribution of the number of defect types and the frequency distribution of each defect type in the battery module to be tested are statistically analyzed. Based on the defect geometric morphology parameters of each candidate defect region in the defect information summary set, geometric morphology parameter distribution analysis is performed on candidate defect regions with the same defect type to generate typical geometric morphology features and geometric parameter variation ranges for each defect type. Spatial distribution analysis is performed on the defect geometric morphology parameters of all candidate defect regions in the defect information summary set to determine the spatial clustering areas and spatial dispersion of different defect types on the surface of the battery module to be tested. Based on the category number distribution of the defect types, the frequency distribution of each defect type, the typical geometric morphological characteristics of each defect type, and the spatial clustering areas of different defect types on the surface of the battery module to be tested, defect distribution statistical description data is generated. Based on the defect distribution statistical description data, the corresponding defect grade scoring threshold is matched from the preset defect grade evaluation standard library to generate the defect grade scoring result of the battery module to be tested; The defect information summary, the defect distribution statistical description data, and the defect grade rating results are combined to form the comprehensive defect detection report.

6. The battery module defect detection method based on deep learning according to claim 3, characterized in that, The method further includes: The intermediate feature data generated by the defect region pre-segmentation network during the candidate region generation process is obtained. The intermediate feature data includes the deep texture fusion feature volume and the generation parameters of the multiple three-dimensional candidate regions. The deep texture fusion feature volume is subjected to feature activation region localization processing to generate a set of spatial coordinates of highly activated response regions in the deep texture fusion feature volume. Based on the set of spatial coordinates, a local three-dimensional surface characterization image patch corresponding to the high activation response region is extracted from the three-dimensional surface characterization image; The local three-dimensional surface characterization image block is subjected to defect detail enhancement processing to generate a high-resolution defect detail image block; The high-resolution defect detail image patch is used as a new training sample to update the weight parameters of the defect region pre-segmentation network; The updated defect region pre-segmentation network is used to perform potential defect region localization processing on the three-dimensional surface characterization image of the battery module to be inspected. Based on the erroneous classification results in the defect category determination results, the classification error is backpropagated to adjust the decision boundary parameters of the defect category classifier.

7. The battery module defect detection method based on deep learning according to claim 1, characterized in that, The method further includes: The process parameter recording data of the production process of the battery module to be tested within a preset historical period is obtained. The process parameter recording data includes welding temperature sequence, welding pressure sequence and liquid injection flow rate sequence. The defect category determination results in the comprehensive defect detection report are subjected to time-series correlation analysis with the process parameter recording data to generate a correlation matrix between defect category and process parameter fluctuation. Based on the correlation matrix, identify the associated process parameter items and process parameter fluctuation thresholds that lead to specific defect types; When the real-time monitoring value of the associated process parameter exceeds the process parameter fluctuation threshold, a process parameter adjustment instruction is generated and fed back to the control unit of the corresponding production process.

8. The battery module defect detection method based on deep learning according to claim 1, characterized in that, The method further includes: Obtain the initial light intensity distribution parameters of the illumination source of the battery module to be tested in the multi-angle imaging device; Brightness uniformity analysis is performed on each battery module surface image unit in the set of appearance images to generate a local brightness deviation distribution map of each battery module surface image unit; Based on the local brightness deviation distribution map, a light source brightness compensation command is generated for the lighting source. The light source brightness compensation command is used to instruct the light source units at different positions to adjust the output light intensity to balance the brightness distribution of the image units on the surface of the battery module. The light source brightness compensation command is sent to the driver controller of the lighting source to trigger the light source parameter adjustment operation; A set of secondary imaging images after adjusting the light source parameters is obtained. The set of secondary imaging images and the set of appearance images are subjected to image quality improvement and fusion processing to generate an optimized set of battery module surface image units. The optimized battery module surface image unit set is used to update the appearance image set, and the viewpoint association mapping process is re-executed to generate an optimized three-dimensional surface characterization image.

9. A battery module defect detection system based on deep learning, characterized in that, include: The processor, communication interface, memory, and communication bus are provided, wherein the processor, communication interface, and memory communicate with each other via the communication bus. The memory is used to store computer programs; the processor is used to execute the computer programs to implement the steps of the deep learning-based battery module defect detection method according to any one of claims 1-8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the deep learning-based battery module defect detection method according to any one of claims 1-8.

Images