Battery pack defect detection method and detection system based on industrial CT

Abstract

This application relates to the fields of automotive batteries, non-destructive testing, and image recognition technology. Specifically, it relates to a method and system for detecting battery pack defects based on industrial CT. The method includes: Step S1, acquiring projection data of the battery pack; based on a digital twin reference, alternating between the projection data domain and the image domain to correct for metal artifacts and reconstruct the entire battery pack, obtaining reconstructed data; the digital twin reference is an ideal state reference body including the geometric information, material information, and regional distribution information of each component within the battery pack. This application can automatically output a detection report with accompanying three-dimensional position.

Description

Technical Field

[0001] This application relates to the fields of automotive batteries, non-destructive trace detection, and image recognition technology. Specifically, it relates to a battery pack defect detection method and system based on industrial CT. Background Technology

[0002] The content in this section provides only background information related to this application and may not constitute prior art.

[0003] A power battery pack is a complex structure formed by multiple processes, including module assembly, busbar welding, and electrolyte injection and encapsulation, consisting of dozens to hundreds of cells. During the manufacturing process of the battery pack, many defects (such as metal foreign objects mixed between the electrodes, welding pores at the welding points of the busbar and the electrode tabs, electrode wrinkles and breaks caused during winding or stacking, electrode alignment deviations, and poor electrolyte wetting) occur inside the cells or are obscured by the casing and module structure. These defects are difficult to detect through visual inspection and electrical performance testing before leaving the factory, but can lead to safety accidents such as capacity decay, internal short circuits, and even thermal runaway during the battery pack's use.

[0004] Industrial CT will be used to perform non-destructive testing on the battery packs that are off the production line. This involves using an X-ray source and detector to collect projection data around the battery pack, reconstructing the three-dimensional volume data inside the battery pack, and then directly subtracting the three-dimensional volume data from a pre-made standard template. Areas with discrepancies will be identified as defects. This solution does not require disassembling the battery pack and has a high degree of automation.

[0005] The battery pack contains large metal components such as busbars and module end plates. When X-rays pass through these metal components, the projection data is severely distorted, resulting in strip-shaped artifacts in the cell area after reconstruction. This not only masks real defects but also creates false defects. At the same time, each cell objectively has legitimate manufacturing tolerances such as fluctuations in electrode thickness and varying winding tightness. When directly subtracted from a fixed standard template, these individual differences that do not constitute defects also manifest as image differences. Furthermore, the electrode layer of a wound cell is a curved surface in space. Directly subtracting from it in the Cartesian coordinates of three-dimensional volume data cannot guarantee that the corresponding positions of the electrode layers are compared with each other.

[0006] Therefore, existing industrial CT scanners cannot accurately determine whether there are defects inside the battery pack. Summary of the Invention

[0007] The summary section of this application is intended to provide a brief overview of the concepts, which will be described in detail in the detailed description section below. This summary section is not intended to identify key or essential features of the claimed technical solutions, nor is it intended to limit the scope of the claimed technical solutions.

[0008] Some embodiments of this application propose a battery pack defect detection method and system based on industrial CT to solve the technical problems mentioned in the background section above.

[0009] As a first aspect of this application, this application provides a battery pack defect detection method based on industrial CT, including: Step S1: Collect projection data of the battery pack; Based on the digital twin reference, perform metal artifact correction and reconstruction on the projection data alternately between the projection data domain and the image domain to obtain the reconstructed data of the entire pack. A digital twin reference is an ideal state reference that includes the geometric information, material information, and regional distribution information of each component within the battery pack; Step S2: Based on the design location of each cell in the digital twin reference, determine the area of ​​each cell in the whole package reconstruction data; For any local area within any cell region, compare it with the corresponding area of ​​the adjacent electrode layer within the same cell to obtain the cross-layer difference; By comparing any local area with the corresponding area of ​​other cells of the same model in the same battery pack, cross-cell differences can be obtained; The cross-layer and cross-cell differences are propagated and aggregated along the sequence direction of the electrode layers, and the aggregated cross-layer and cross-cell differences are fused to obtain a defect confidence field covering the entire package reconstruction data. Step S3: Perform three-dimensional aggregation of the defect confidence field to obtain candidate defects, determine the three-dimensional location and risk level of each candidate defect, and generate a battery pack inspection report.

[0010] As a second aspect of this application, this application provides a battery pack defect detection system based on industrial CT, comprising: The scanning device, arranged at the scanning station, includes a radiation source and detector that move around the battery pack along a spiral trajectory, and a positioning fixture that keeps the battery pack stationary during the scanning process. Data storage server, which houses the digital twin reference and the good product reference memory; An algorithm server is used to run the aforementioned battery pack defect detection method based on industrial CT.

[0011] The technical solutions of this application have at least the following advantages and beneficial effects: This application utilizes a digital twin reference body to span the entire process of acquisition, reconstruction, comparison, and judgment, thereby simultaneously reducing both missed detections and false alarms in industrial CT inspection of battery packs, and ensuring that the inspection results are interpretable and executable. Based on the regional distribution information of the digital twin reference body, the penetration conditions are switched in different segments, taking into account both the penetration capability of dense metal areas and the low-contrast detail resolution of the cell body area. The ideal projection shape obtained by forward projection of the metal geometric model is used as a constraint marker to repair the contaminated section, and the alternating iteration between the projection data domain and the image domain suppresses metal artifacts at the source of reconstruction rather than afterward, thus preserving the details of the cell area. The surface unfolding transformation flattens the curved electrode layer into a stacked image sequence under a unified layer sequence coordinate, so that the corresponding positions across layers and across cells can be directly indexed, and minor defects become comparable and locatable. The differences between layers and across cells are double-verified through propagation aggregation and gating fusion. Isolated noise is diluted during propagation, and legitimate manufacturing tolerances and common process deviations of the whole package are naturally suppressed by the gating mechanism. Only real defects that simultaneously destroy the consistency between layers and between cells are judged as high-confidence anomalies. Moreover, the propagation weight and judgment threshold are self-learned and calibrated by good product data and updated online with the production line status, without the need for repeated manual parameter tuning, and automatically outputs a detection report with three-dimensional position. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of a scenario where a scanning device performs a spiral CT scan on a battery pack.

[0013] Figure 2 This is a schematic diagram of the surface unfolding transformation of the battery cell region.

[0014] Figure 3 This is a schematic diagram illustrating the calculation, propagation aggregation, and gating fusion of cross-layer and cross-cell differences.

[0015] Figure 4 This is a schematic diagram illustrating the marking and repair of contaminated sections in metal artifact correction based on a digital twin reference.

[0016] Figure 5 This is a schematic diagram of the three-dimensional positioning view in the inspection report.

[0017] Figure 6 This is a schematic diagram of the components of a battery pack defect detection system based on industrial CT.

[0018] Figure 7 This is a schematic diagram of the process of correcting and reconstructing metal artifacts based on a digital twin reference. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments. The same reference numerals in the accompanying drawings represent the same components. It should be noted that the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the described embodiments of this application without creative effort are within the scope of protection of this application.

[0020] Compared to the embodiments shown in the accompanying drawings, feasible embodiments within the scope of this application may have fewer components, other components not shown in the drawings, different components, differently arranged components, or components with different connections, etc. Furthermore, two or more components in the drawings may be implemented in a single component, or a single component shown in the drawings may be implemented as multiple separate components.

[0021] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms "first," "second," and similar terms used in this specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, "one" or "a similar" does not necessarily indicate a quantity limitation. Terms such as "upper," "lower," etc., are used only to indicate relative positional relationships, and these relative positional relationships may change accordingly when the absolute position of the described object changes.

[0022] The technical concept of this application is as follows: The principle of industrial CT inspection of battery packs is as follows: X-rays emitted from a radiation source penetrate the battery pack and are received by a detector, forming frames of projection data. After acquiring projection data from a sufficient number of angles, a reconstruction algorithm can reconstruct the three-dimensional volumetric data of the battery pack's interior. The value of each voxel in the volumetric data reflects the material's attenuation ability to X-rays at that location. Using this volumetric data, the electrode layers, separators, tabs, and welded areas inside the battery cells can be observed, allowing for the inspection of internal defects such as foreign metal objects, weld porosity, and electrode wrinkles without disassembling the battery pack.

[0023] The battery pack contains large metal components, which attenuate X-rays significantly when passing through them. This causes severe distortion in the projected data in the corresponding sections, resulting in strip-shaped artifacts running through the cell area after reconstruction. These artifacts can either mask real defects, leading to missed detections, or they can be mistaken for defects, causing false alarms. If ordinary interpolation is used to directly fill in the contaminated projected data, the filled content lacks physical basis, and it is equally difficult to distinguish the filled structure from the real structure.

[0024] To this end, this application pre-constructs a digital twin reference body that includes the geometric information, material information, and regional distribution information of each component within the battery pack. During the imaging stage, the ideal projected shape of the metal is projected frame-by-frame using the geometric model of the metal components within the digital twin reference body. Based on this, sections contaminated by metal are located and repaired in the projection data. The process iterates alternately between the projection data domain and the image domain to suppress metal artifacts at their source. During the comparison stage, each cell region is first located based on the digital twin reference body. A surface unfolding transformation is performed on each cell, flattening the curved electrode layers layer by layer into a stacked image sequence, allowing direct indexing of corresponding positions between layers and between cells.

[0025] Subsequently, cross-layer and cross-cell differences are calculated simultaneously for each local area, propagated and aggregated along the sequence direction, and fused using a gating method that only judges an anomaly when both types of differences exceed the limit simultaneously, so that isolated noise and individual differences are naturally suppressed; finally, the defect confidence field obtained by fusion is three-dimensionally aggregated, and a test report that can directly guide production line handling is output.

[0026] The implementation scenarios for the following embodiments are set as follows: After the battery pack comes off the production line, it is conveyed to the scanning station by a conveyor line and fixed by a positioning fixture to keep it stationary during the scanning process.

[0027] The X-ray source and detector are mounted on a spiral scanning mechanism surrounding the battery pack. During scanning, the X-ray source and detector rotate around the battery pack and translate along the length of the battery pack, forming a spiral trajectory and acquiring projection data frame by frame.

[0028] The X-ray source, detector, and their motion mechanism are existing technologies, and their specific structures will not be further described in this application; the key to this application lies in the following processing procedure performed on the acquired projection data and reconstructed volume data.

[0029] refer to Figure 1 Example 1: A battery pack defect detection method based on industrial CT.

[0030] The method in this embodiment includes the following steps: Step S1: The X-ray source and detector move around the battery pack along a spiral trajectory to collect the projection data of the battery pack. Based on the digital twin reference, the projection data is alternately corrected for metal artifacts and reconstructed between the projection data domain and the image domain to obtain the whole package reconstructed data. A digital twin reference is an ideal state reference that includes the geometric information, material information, and regional distribution information of each component within the battery pack; Step S2: Based on the design location of each cell in the digital twin reference, determine the area of ​​each cell in the whole package reconstruction data; For any local area within any cell region, compare it with the corresponding area of ​​the adjacent electrode layer within the same cell to obtain the cross-layer difference; By comparing any local area with the corresponding area of ​​other cells of the same model in the same battery pack, cross-cell differences can be obtained; The cross-layer and cross-cell differences are propagated and aggregated along the sequence direction of the electrode layers, and the aggregated cross-layer and cross-cell differences are fused to obtain a defect confidence field covering the entire package reconstruction data. Step S3: Perform three-dimensional aggregation of the defect confidence field to obtain candidate defects, determine the three-dimensional location and risk level of each candidate defect, and generate a battery pack inspection report.

[0031] The above outlines the general steps of this method and the general function of each step.

[0032] For ease of understanding, the following explanation of the digital twin reference body is provided beforehand: The digital twin reference body is pre-built, and the geometric information of the digital twin reference body is given by the CAD model of the battery pack and its internal components (shell, busbar, module end plate, cell, fastener, etc.). Material information includes the material type of each component and its attenuation characteristics for X-rays; The regional distribution information is the spatial distribution of various components along the scanning axis and the cross-sectional direction.

[0033] Thus, a digital twin reference can be understood as a digital prototype of this battery pack model under ideal conditions, telling subsequent steps what materials and structures should be used at each location. A digital twin reference is built once for each battery pack model, and shared by all battery packs of the same model.

[0034] The specific implementation methods for each step are explained below: Step S1 includes the following steps: Step S11: Place the battery pack at the scanning station. The X-ray source and detector move around the battery pack along a spiral trajectory to collect projection data and record the spatial pose of the X-ray source and detector frame by frame. Specifically, the spiral trajectory allows a single continuous scan to cover the entire length of the battery pack, avoiding splicing errors caused by segmented scanning.

[0035] Recording the spatial pose of the X-ray source and detector frame by frame means synchronously recording the position and orientation of the X-ray source focus and detector panel in a unified coordinate system when the projection data of each frame is acquired. This pose record is used in step S14 to determine the projection geometry of each frame, which is the basis for subsequent frame-by-frame forward projection and frame-by-frame repair.

[0036] During the acquisition process, the spiral path is divided into several scanning segments based on the regional distribution information in the digital twin reference body, and the penetration conditions of the X-ray source and the acquisition conditions of the detector are switched according to the scanning segments. The regional distribution information includes the busbar area, module end board area, and cell body area.

[0037] Specifically, the busbar area and module end plate area have dense metal, so a higher tube voltage and a hardened filter are used to ensure the penetration ability of the radiation through the metal. The cell body area is mainly made of lightweight materials, and uses a lower tube voltage, a higher detector gain, and a longer integration time to ensure that low-contrast details can be resolved.

[0038] For example, the busbar area voltage is about 450 kV, and the cell body area voltage is about 225 kV. The specific values ​​are determined based on the battery pack size and materials and are preset values.

[0039] The projection data is then subjected to background correction, gain correction, and bad pixel repair in sequence to obtain the corrected projection data.

[0040] Background correction, gain correction, and bad pixel repair are routine preprocessing steps for industrial CT detectors and are existing technologies. This application will not further describe their specific processes.

[0041] Step S12: Perform initial reconstruction on the projection data to obtain a preliminary reconstructed volume.

[0042] Specifically, the initial reconstruction can be achieved using analytical reconstruction algorithms commonly used in spiral cone-beam CT. These algorithms are existing technologies, and their specific processes will not be further described in this application. Reconstruction refers to converting projection data into three-dimensional CT image data.

[0043] It should be noted that obvious strip-shaped artifacts still exist around the metal components in the preliminary reconstruction. Their function is not to be used directly for defect determination, but to provide the measured structural profile for spatial registration in step S13.

[0044] Step S13: Spatial registration is performed between the digital twin reference body and the preliminary reconstructed body to obtain the spatial transformation relationship between the digital twin reference body and the preliminary reconstructed body.

[0045] Spatial registration is performed in two stages: The first stage uses the battery pack's housing outline and / or mounting holes as features for coarse registration, eliminating overall translational and rotational deviations. The second stage involves flexible fine registration based on coarse registration, absorbing local deformations caused by manufacturing tolerances and assembly deviations, to obtain the aforementioned spatial transformation relationship.

[0046] Specifically, coarse registration is rigid registration: The outer contour surface of the battery pack casing and the center of the mounting holes are extracted from the initial reconstruction and matched with corresponding features in the digital twin reference body to solve for an overall rotation matrix and translation vector. The casing contour and mounting holes are chosen as features because they are large in size, geometrically stable, and less affected by metal artifacts, allowing for reliable extraction even if the initial reconstruction quality is not high.

[0047] Based on the coarse registration results, elastic fine registration uses elastic transformations such as B-spline free deformation to absorb local deformation.

[0048] Both coarse registration and flexible registration algorithms are existing technologies, and their specific processes will not be further described in this application.

[0049] The reason for the two-stage process is that the actual battery pack has both overall placement deviations relative to the design state and local deformations caused by manufacturing tolerances and assembly deviations. If only rigid registration is performed, the residual local deformation will cause the ideal projection shape obtained in step S14 to be misaligned with the actual metal projection, and the contaminated section will be inaccurate. If flexible registration is performed directly without coarse registration, the flexible registration is prone to falling into incorrect local correspondences when the initial deviation is too large.

[0050] Step S14: Based on the spatial transformation relationship, transform the geometric model marked as metal material in the digital twin reference body to the coordinate space where the initial reconstructed body is located; For the transformed geometric model, forward projection is performed frame by frame according to the projection geometry determined by the spatial pose recorded frame by frame to obtain the ideal projection shape of the metal component. Specifically, forward projection involves connecting each ray from the focal point of the ray source to the detector pixel by pixel according to the actual pose of the ray source and detector when the frame is acquired. This process calculates the intersection of each ray with the metal geometric model, and determines where the metal should appear on the projection of this frame, which pixels it should cover, and what the corresponding penetration thickness should be if the battery pack is completely consistent with the design. This is the ideal projection shape of the frame.

[0051] Based on the ideal projection shape, the contaminated sections of the metal components through which the rays pass are marked frame by frame in the projection data. The contaminated sections are then repaired to obtain the repaired projection data.

[0052] The remediation methods for the contaminated sections are as follows: Using uncontaminated data adjacent to the contaminated section within the same frame as a reference, and constrained by the ideal projection shape obtained by forward projection of the metal component in the digital twin reference, the data in the contaminated section is interpolated and reconstructed along the detector direction.

[0053] refer to Figure 4Specifically, when marking contaminated sections, a certain number of pixels can be added outside the boundary of the ideal projection shape as a safety margin. The amount of addition is a preset value, such as 2 to 5 pixels, to cover registration residual errors and scattering smearing. In other words, the contaminated section is the image extended from the ideal projection shape.

[0054] The marking of the contaminated section can be expressed mathematically. Within any frame, let β be the pixel position along the detector direction. Then, the mask M of the contaminated section is determined according to equation (S14-1): When min{|β-β′|:[Φ(W)](β′)>0}≤ν, M(β)=1; In other cases, M(β) = 0; (S14-1) Where β represents the pixel position along the detector direction within the frame; M is the mask for the contaminated section, M(β) represents the value of the mask at pixel β in the frame, M(β)=1 indicates that pixel β belongs to the contaminated section, M(β)=0 indicates that pixel β belongs to uncontaminated data; Φ(·) represents the forward projection operator executed according to the projection geometry determined by the spatial pose recorded frame by frame; W represents the metal backfill mask, W takes 1 at the voxels covered by the geometric model of the metal material in the digital twin reference after spatial transformation and 0 at the other voxels; [Φ(W)](β) represents the value of the mask after forward projection of W in the frame. The value at pixel β has the physical meaning of the penetration thickness of the ray corresponding to that pixel through the metal component. The distribution of [Φ(W)](β) along the detector direction is the ideal projection shape of the frame. [Φ(W)](β)>0 indicates that pixel β is located within the ideal projection shape. β′ represents the position of another pixel in the frame. |β-β′| represents the distance between the two pixel positions. min{·} represents taking the minimum value in the set, that is, min{|β-β′|:[Φ(W)](β′)>0} is the shortest distance from pixel β to the ideal projection shape. ν is the outward expansion amount, and ν is a preset value, for example, 2 to 5 pixels.

[0055] Equation (S14-1) means that all pixels within the ideal projection shape and the outer range of no more than ν pixels from the ideal projection shape together constitute the contaminated section.

[0056] During the repair process, the starting and ending points of the interpolation are taken from the uncontaminated data adjacent to the contaminated section on both sides, and the shape of the interpolation curve within the section is constrained by the ideal projection shape.

[0057] The ideal projection shape gives the trend of the thickness of the metal penetration within the section, and the interpolation result should be consistent with this trend, rather than a simple straight line connection.

[0058] The interpolation reconstruction of the contaminated section can be expressed by a mathematical formula: Let ζ and ζ′ be the positions of the uncontaminated pixels immediately to the left and right of the contaminated segment within the frame, respectively, which are the start and end points of the interpolation. Then, the restored projection data Q(0) within the contaminated segment is calculated according to formula (S14-2): When M(β)=1, Q(0) (β) =[(ζ′-β)·Q(0) (ζ) +(β-ζ)·Q(0)( ζ′) ] / (ζ′-ζ)+κ·[Φ(W)](β), (S14-2); Where Q(0) represents the repaired projection data obtained in step S14, Q(0) (β) This indicates the value at pixel β in that frame; when M(β)=0, Q(0) (β) The original value of the corrected projection data at that location is taken directly, meaning that the uncontaminated data is not modified; Both ζ and ζ′ are uncontaminated data, therefore Q(0) (ζ) With Q(0) (ζ′) That is, the two reference values ​​on which the interpolation is based; the first term in equation (S14-2) [(ζ′-β)·Q(0)] (ζ) +(β-ζ)·Q(0) (ζ′) ] / (ζ′-ζ) is a linear background term determined by the reference values ​​at both ends, representing the smooth trend of the projection value in this section when there are no metal components; the second term κ·[Φ(W)](β) is an ideal projection shape constraint term, which makes the interpolation curve fluctuate within the section as the metal penetration thickness [Φ(W)](β) increases or decreases, that is, the trend of the metal thinning here is written into the repair result; κ is the equivalent attenuation amplitude coefficient of the metal component, κ is a non-negative real number, which can be determined based on the attenuation characteristics given by the material information of the metal component in the digital twin reference, or it can be obtained by fitting the measured data adjacent to the edge of the ideal projection shape in the same frame. Since ζ and ζ′ are located outside the outer range, [Φ(W)](ζ)=[Φ(W)](ζ′)=0, so the values ​​of Equation (S14-2) are continuous and naturally connected with the uncontaminated data at both ends of the contaminated section, and will not produce step-like false signals at the metal edge.

[0059] The reason for this interpolation method is that ordinary interpolation only guarantees numerical continuity but not physical rationality, and it is easy to create step-like false signals at the metal edges. Using the ideal projection shape as a constraint is equivalent to telling the interpolating metal to become thinner here, so that the repaired data has a physical basis and the subsequent reconstruction will not fabricate or erase the structure.

[0060] Step S15: Reconstruct using the repaired projection data, and iteratively correct the repair results of the contaminated section by alternating between the projection data domain and the image domain.

[0061] refer to Figure 7 Specifically, the iterative process is as follows: A new reconstructed volume is obtained using the repaired projection data; this is the image domain side. In the new reconstructed volume, the stripe artifacts around the metal have been weakened, allowing for a more accurate estimation of the true boundary of the metal component and its surrounding structure. The updated structure is then projected back into the projection data domain, and the repair content of the contaminated sections in each frame is corrected accordingly; this is the projection data domain side. Reconstruction and correction are repeated. In each iteration, the image domain provides clearer structural information, while the projection data domain provides more faithful constraints. The two mutually correct each other, and the repair result gradually approaches the true value.

[0062] During the alternating iteration process, the intensity of residual strip artifacts corresponding to the cell body region in the image domain is continuously monitored. When the intensity of residual strip artifacts converges and tends to stabilize, the alternating iteration is terminated.

[0063] Specifically, the intensity of residual strip artifacts can be measured by the high-frequency component energy along the artifact characteristic direction within the cell body region. When the relative change of this measure in two adjacent iterations is less than a preset convergence threshold (e.g., 1% to 3%) and remains so for several consecutive iterations (e.g., 2 to 3 iterations), convergence is considered to be stable. The reason for using the cell body region as the monitoring object is as follows: Defect detection occurs in the cell area, and the purpose of alternating iterations is to clean the cell area. The imaging quality of the metal area itself is not the focus of this method, and using it as the monitoring object would waste iterations.

[0064] After the iteration terminates, the geometric model of the metal material in the digital twin reference body is backfilled into the corresponding position of the reconstruction result to obtain the complete reconstructed body data.

[0065] The above-mentioned alternating iteration and backfilling process can be represented by a mathematical formula.

[0066] In the alternating iteration, the reconstruction in the j-th round is performed according to formula (S15-1), and the modification of the remediation content of the contaminated section is performed according to formula (S15-2): Z(j)=Ψ(Q(j)) (S15-1; Q(j+1)=M⊙Π(Φ(Z(j)))+(1-M)⊙Q(0) (S15-2); Where j represents the round of alternating iteration, j=0,1,2,... in ascending order; Q(j) represents the repaired projection data of the j-th round, Q(0) is the repaired projection data obtained in step S14; Ψ(·) represents the reconstruction operator, that is, the operation of reconstructing the reconstructed body from the projection data; Z(j) represents the reconstructed body obtained in the j-th round; Φ(·) represents the forward projection operator executed according to the projection geometry determined by the spatial pose recorded frame by frame; Π(·) represents the repair operator that performs interpolation reconstruction of the contaminated section along the detector direction with the ideal projection shape as a constraint; M represents the mask of the contaminated section, M takes 1 in the contaminated section and 0 in other positions; ⊙ represents element-wise multiplication at corresponding positions; the (1-M)⊙Q(0) term in equation (S15-2) ensures that the uncontaminated data retains the original measured value without being modified throughout the iteration process.

[0067] The combined meaning of equations (S15-1) and (S15-2) is as follows: In each round, the reconstructed volume is first obtained from the current projection data, which is the image domain side; then the reconstructed volume is projected back into the projection data domain by the forward projection operator Φ(·), and the repaired content in the contaminated section is corrected by the repair operator Π(·), while the uncontaminated section maintains its original value, which is the projection data domain side; and so on.

[0068] The termination criterion for alternating iterations is performed according to equation (S15-3): |σ(j)-σ(j-1)| / σ(j-1)<ε (S15-3); Where σ(j) represents the intensity of the residual strip artifact corresponding to the cell body region in the reconstructed body Z(j) of the j-th round, that is, the high-frequency component energy along the artifact characteristic direction in the cell body region as mentioned above; |·| represents taking the absolute value; ε is the preset convergence threshold mentioned above, for example, 1% to 3%. When equation (S15-3) holds for several rounds (e.g., 2 to 3 rounds), it is determined that the intensity of the residual strip artifact has converged and tended to be stable, the alternating iteration is terminated, and the round at the time of termination is recorded as η.

[0069] The backfilling of the metal geometry model after the iteration terminates shall be performed according to formula (S15-4): Z*=(1-W)⊙Z(η)+W⊙U(S15-4); Where Z(η) represents the reconstructed body corresponding to the terminating round η; W represents the metal backfill mask, where W is set to 1 at the voxels covered by the geometric model of the metal material in the digital twin reference body after spatial transformation and mapping, and 0 at the other voxels; U represents the metal value field that is assigned a certain value according to the material of the metal component; Z* is the data of the entire reconstructed body obtained after backfilling.

[0070] Specifically, backfilling refers to assigning a definite value to the space occupied by metal components according to the shape and material of the geometric model, replacing the content that is still not completely reliable after repair and reconstruction.

[0071] The metal components themselves (busbars, module end plates, etc.) are not the objects of defect detection, and their internal grayscale does not need to be preserved. After backfilling, the metal boundaries are clear and the positions are accurate, which not only avoids residual artifacts from seeping from the metal area to the cell area, but also provides a clean spatial reference for step S2 to locate the cell area according to the design position.

[0072] Thus, in the whole-pack reconstruction data output in step S1, metal artifacts are suppressed while details of the cell area are preserved: the ideal state given by the digital twin reference and the measured projection data are mutually verified in the alternating correction between the projection data domain and the image domain, and the reconstruction quality is guaranteed from the source.

[0073] A pixel is the smallest photosensitive unit on the detector plane, belonging to the side of the projected data (two-dimensional). X-rays pass through the battery pack and hit the detector, which is a panel made up of thousands of tiny photosensitive grids. Each grid records an attenuation reading, and one grid is a pixel. You can think of each frame of projected data as a grayscale photograph, and a pixel is a single pixel in that photograph.

[0074] A voxel is the smallest volumetric unit in 3D reconstructed volume data, belonging to the reconstructed 3D space side; it's essentially extending 2D pixels to 3D volumetric pixels. The reconstruction algorithm restores projection data from different angles into a 3D data block (the entire reconstructed volume data).

[0075] The relationship between pixels and voxels is as follows: pixels are units at the acquisition end (on the detector), while voxels are units at the reconstruction end (in 3D data). The former is transformed into the latter through a reconstruction algorithm.

[0076] The central surface (which will appear later) is not a data unit, but a geometric reference surface: the electrode layer of a wound battery cell itself has a thickness (the electrode layer plus the separator is about a few tenths of a millimeter), which is represented as a curved "thin shell" in three-dimensional volume data. The central surface is the imaginary surface in the middle along the thickness direction of this thin shell, just like a piece of paper has two sides, and the central surface is the surface in the middle of the paper core. Its function is to provide a framework for the unfolding and transformation of the surface.

[0077] Step S2: Based on the design location of each cell in the digital twin reference, determine the area of ​​each cell in the whole package reconstruction data; For any local area within any cell region, compare it with the corresponding area of ​​the adjacent electrode layer within the same cell to obtain the cross-layer difference; By comparing any local area with the corresponding area of ​​other cells of the same model in the same battery pack, cross-cell differences can be obtained; The cross-layer and cross-cell differences are propagated and aggregated along the sequence direction of the electrode layers, and the aggregated cross-layer and cross-cell differences are fused to obtain a defect confidence field covering the entire package reconstruction data. refer to Figure 2 Next, we will explain the specific implementation method of step S2.

[0078] Based on the design location of each cell in the digital twin reference, the area of ​​each cell is determined in the overall reconstructed data.

[0079] Specifically, using the spatial transformation relationship obtained in step S13, the design bounding box of each cell in the digital twin reference body is transformed to the coordinate space of the whole package reconstruction data, and the actual boundary of the cell is refined according to grayscale and structural features within the bounding box to obtain each cell region; each cell region is assigned a unique cell number, and its cell model is recorded according to the digital twin reference body.

[0080] Before calculating the differences between layers and between cells, a surface unfolding transformation is performed on each cell region.

[0081] In a wound cell, the electrode layers are coiled around the core, and the same layer of electrode is a curved surface in space. When comparing adjacent positions directly in the Cartesian coordinates of three-dimensional volume data, what is obtained is often not the adjacent content of the same electrode layer, but a mixture of content that spans electrodes, separators, and even adjacent layers.

[0082] The surface unfolding transformation can be understood as flattening a roll of paper: after flattening, each layer of paper becomes a flat image, and the layers are naturally aligned. The kth layer, the nth row, and the nth column become the coordinate language common to all battery cells.

[0083] The surface unfolding transformation is to extract the central surface of each electrode layer within the cell region.

[0084] The extraction method for the central surface is as follows: For wound cells, directional structural reinforcement is first applied to the cell area. Then, starting from the center area of ​​the core, the center position of each electrode layer is traced outward layer by layer along the direction of the electrode layer structure to obtain a continuous center surface of each electrode layer. Specifically, directional structure enhancement processing is used to highlight layered structures and suppress noise. It can be implemented by layered structure enhancement filtering based on structural tensors or Hessian matrices. This type of filtering is existing technology, and its specific process will not be described further in this application.

[0085] For laminated cells, the center plane of each electrode is fitted layer by layer along the lamination direction to form the center surface.

[0086] Specifically, the electrodes of the laminated battery cell are approximately planar, and planar fitting can be performed layer by layer along the lamination direction. The planar fitting method is existing technology, and its specific process will not be described further in this application.

[0087] For each central surface, an in-plane two-dimensional parameterization is established. The entire reconstructed volume data is resampled within a preset thickness range on both sides of the central surface along the normal direction. Each pole piece is flattened into a pseudo-planar image. The pseudo-planar images of all pole pieces are arranged in sequence to form a stacked image sequence in the sequence coordinate system.

[0088] Specifically, in-plane two-dimensional parameterization refers to establishing two orthogonal in-plane coordinates on the central curved surface (for wound cells, two coordinates can be taken along the winding direction and along the cell height direction), so that each point on the curved surface obtains a unique two-dimensional coordinate.

[0089] The preset thickness range is slightly larger than the nominal thickness of a single-layer electrode and diaphragm, for example, 1.2 to 1.5 times the nominal thickness, which ensures that the sampling covers the entire layer without mixing in too much of the adjacent layers.

[0090] In a pseudo-planar image, row and column coordinates are in-plane two-dimensional coordinates, and pixel values ​​are grayscale values ​​obtained by resampling along the normal direction. A sequence of stacked images can be denoted as {Ik}, where k is the layer sequence number of the pole piece, increasing sequentially from the center of the core outwards.

[0091] During resampling, the correspondence between each sampling point in the stacked image sequence and the coordinates of the entire reconstructed volume data is recorded synchronously to form a forward and reverse coordinate mapping table.

[0092] Specifically, forward mapping retrieves the layer number and in-plane coordinates from the 3D coordinates of the reconstructed volume data in the stacked image sequence, while reverse mapping retrieves the 3D coordinates from the layer number and in-plane coordinates. The forward and reverse coordinate mapping tables can be understood as a bidirectional dictionary between two sets of coordinates. Difference calculations are performed on the unfolded stacked image sequence, but the final location and reporting of defects must return to the actual 3D space of the battery pack. The mapping table ensures lossless travel between the two spaces.

[0093] The differences between layers and between cells are calculated on the stacked image sequence. The fused results are mapped back to the coordinate space of the whole package reconstructed data through a forward and reverse coordinate mapping table to form the defect confidence field.

[0094] refer to Figure 3 In step S2, the calculation, propagation aggregation, and fusion of cross-layer differences and cross-cell differences include the following steps: Step S21: Select a local region on any layer of the stacked image sequence using a sliding window method, and extract the gray-level distribution features and local structure description of the local region.

[0095] Specifically, the window size and step size of the sliding window are preset values, such as a window size of 16×16 pixels and a step size of 8 pixels. Adjacent windows partially overlap to avoid defects falling exactly on the window boundary and being cut off.

[0096] Gray-level distribution features represent the statistical form of gray levels within a local area, such as gray-level mean, variance, and histogram; local structure descriptions represent the orientation and texture of structures within a local area, such as gradient direction statistics or texture descriptors.

[0097] The specific calculation methods for grayscale distribution characteristics and local structure description are existing technologies, and this application will not further describe their specific processes.

[0098] Step S22: Search for the most similar reference neighborhood near the corresponding positions of several adjacent layers, compare the differences between the local region and the reference neighborhood in terms of gray-scale distribution features and local structure description, and obtain the cross-layer difference of the local region.

[0099] Specifically, several adjacent layers can be selected, for example, one or two layers above and below each other; Near-location retrieval refers to finding the most structurally similar neighborhood within a preset retrieval radius (e.g., 3-5 pixels) centered on the coordinates within the same plane, rather than rigidly using the same coordinates.

[0100] In this way, there are legitimate small misalignments between the patterns of adjacent electrode layers, allowing small-scale searches to absorb these misalignments and avoid mistaking the misalignment itself for a difference; at the same time, an upper limit is set on the search radius to prevent searching for similarities in distant areas and missing the real differences.

[0101] Specifically, within a preset retrieval radius, the similarity between the local region and each candidate neighborhood is calculated one by one. The candidate neighborhood is a region within the preset retrieval radius that has the same size as the local region; the candidate neighborhood with the highest similarity is the reference neighborhood.

[0102] The difference can be taken as a weighted sum of the differences in grayscale distribution features and the differences in local structure description, and normalized to [0,1], with the weights being preset values. Cross-layer differences characterize whether a point is abnormal compared to the layers above and below it.

[0103] The difference in grayscale distribution features is determined by the similarity of grayscale distribution features between the local region and the reference neighborhood. The higher the similarity, the smaller the difference. For example, the difference can be calculated by subtracting the normalized similarity from 1. The difference in local structural description is determined similarly by the similarity of local structural description. In other words, the similarity of grayscale distribution features and the similarity of local structural description are calculated for the retrieved reference neighborhood and the local region, respectively, and then the corresponding difference is obtained.

[0104] Step S23: In the stacked image sequence of other cells of the same model in the same battery pack, the corresponding reference area is indexed according to the layer number and in-plane position. The differences between the local area and the corresponding reference area in gray-scale distribution features and local structure description are compared to obtain the cross-cell difference of the local area.

[0105] Specifically, because all battery cells are unfolded into a unified layer sequence coordinate system, the index of the k-th layer and in-plane position p is universal across all cells, allowing for direct indexing across cells without further registration. Several corresponding reference areas from other cells of the same model can be compared separately, with the cross-cell difference calculated as the median of the comparison results. This prevents defects in individual reference cells from contaminating the comparison benchmark. The cross-cell difference characterizes whether this location is abnormal compared to the same position in other cells within the same package.

[0106] For local areas with insufficient references, the most similar sample is retrieved from the good product reference memory as a supplementary reference; The good product reference memory is a collection of local structural descriptions of stacked image sequences of good product battery packs that have been re-inspected and confirmed during the deployment phase.

[0107] Specifically, insufficient reference refers to a situation where the number of usable cells of the same model within the same battery pack is insufficient (e.g., in a small-capacity battery pack), or where the corresponding position is exactly covered by the position of interest on most reference cells. The good product reference memory uses the cell model as an index, and during retrieval, it selects the most similar sample based on the layer number, in-plane position, and similarity of local structural descriptions. The construction and maintenance method of the good product reference memory is described in Example 2.

[0108] Step S24: After completing the layer-by-layer calculation, proceed from the inner layer to the outer layer and then from the outer layer to the inner layer along the layer sequence direction. In this way, merge the cross-layer difference and cross-cell difference of each layer with the aggregation results of the adjacent aggregated layers according to the propagation weight to complete the propagation aggregation. The propagation weights are obtained by self-learning calibration based on the interlayer attenuation law of noise differences in the stacked image sequence of good battery packs.

[0109] Specifically, propagation aggregation is performed on cross-layer differences and cross-cell differences using the following formulas: Ek(p)=(1-γk)·dk(p)+γk·Ek′(p*) (1); Where, k represents the layer sequence number of the electrode layer; p represents the in-plane position of the local region on the k-th layer; dk(p) represents the cross-layer difference obtained by step S22 or the cross-cell difference obtained by step S23 for the local region at position p of the k-th layer; k′ represents the layer sequence number of the adjacent aggregated layer, k′=k-1 when aggregated from the inner layer to the outer layer, and k′=k+1 when aggregated from the outer layer to the inner layer; p* represents the position in the k′-th layer corresponding to position p; Ek′(p*) represents the aggregation result of the k′-th layer at position p*; Ek(p) represents the aggregation result of the k-th layer at position p; γk is the propagation weight, with a value range of [0,1], obtained by the self-learning calibration described in step S24.

[0110] Equation (1) is a recursion of current observation + neighboring layer memory: the aggregation result of the current layer is obtained by weighting the difference dk(p) of the layer itself and the aggregation result Ek′(p*) of the adjacent aggregated layers. The larger γk is, the stronger the influence of the neighboring layer evidence. The reason for propagation aggregation is that: real defects (such as foreign object indentations, electrode wrinkles) often affect adjacent layers along the sequence direction, and the differences of each layer appear in the same direction; while noise is independent between layers. The recursion makes the differences that appear consistently across layers accumulate and strengthen layer by layer, while isolated noise differences are diluted in the propagation. The reason for scanning from the inner layer to the outer layer and then from the outer layer to the inner layer is that: unidirectional propagation only allows the evidence of the inner neighboring layers to support the current layer, and the layer located on the outer edge of the defect does not get the outer evidence; after scanning in both directions, each layer absorbs the evidence from both sides at the same time, and the aggregation result is not biased to either side.

[0111] For example: Suppose that the aggregation from the inner layer to the outer layer reaches a position p in a certain plane of the k=12th layer, where the inter-layer difference d12(p) = 0.80, the aggregation result at the corresponding position in the 11th layer E11(p*) = 0.50, and the propagation weight γ12 = 0.30 (the calibration process is shown in Example 2), then according to formula (1): E12(p) = (1-0.30)×0.80 + 0.30×0.50 = 0.71. The inter-cell difference at the same position is aggregated independently according to the same formula: Suppose that the inter-cell difference d12(p) = 0.75, and the inter-cell aggregation result at the corresponding position in the 11th layer E11(p*) = 0.40, then the aggregation result is (1-0.30)×0.75 + 0.30×0.40 = 0.645.

[0112] Step S25: When the cross-layer difference after aggregation of any local region and the cross-cell difference after aggregation both exceed their respective judgment thresholds, it is judged as a high-confidence anomaly. Local areas where a single difference exceeds the judgment threshold will be downweighted. All high-confidence anomalies are used as gated fusion results; The judgment threshold is adaptively calibrated on the stacked image sequence of good battery packs according to the preset false alarm control level; The result of gating fusion is mapped back to the coordinate space of the whole package reconstructed volume data through a forward and reverse coordinate mapping table to obtain the defect confidence field.

[0113] Let GL(k,p) be the result of cross-layer difference propagation aggregation according to equation (1), and GC(k,p) be the result of cross-cell difference propagation aggregation. That is, GL(k,p) represents the cross-layer difference after aggregation at position p in the k-th layer, and GC(k,p) represents the cross-cell difference after aggregation at position p in the k-th layer. Then, the confidence value F(k,p) obtained by gating fusion of the local region at position p in the k-th layer is performed according to equation (S25-1): When GL(k,p)≥θL and GC(k,p)≥θC: F(k,p)=min(GL(k,p) / θL, GC(k,p) / θC); When only one of the two conditions, GL(k,p)≥θL and GC(k,p)≥θC, is true: F(k,p)=λ·min(GL(k,p) / θL, GC(k,p) / θC); When neither of the conditions GL(k,p)≥θL nor GC(k,p)≥θC holds true, F(k,p)=0 (S25-1). Where θL represents the threshold for determining cross-layer differences after aggregation; θC represents the threshold for determining cross-cell differences after aggregation; min(·,·) means taking the smaller of the two terms in parentheses; GL(k,p) / θL and GC(k,p) / θC are two normalized aggregation differences; λ is the weighting coefficient, which is a preset value with a range of (0,1), for example, 0.3.

[0114] When both conditions are met, i.e., the differences between the two types of aggregations both exceed their respective judgment thresholds, F(k,p) is not less than 1, which corresponds to the judgment of a high confidence anomaly; when only one of the two conditions is met, F(k,p) is not greater than λ, i.e., it is downweighted. When neither of the two conditions is met, F(k,p) takes the value 0, which is considered normal.

[0115] Based on the above two conditions, the value of F(k,p) can be obtained, and the defect confidence field can be calculated based on the value of F(k,p).

[0116] The defect confidence field is executed according to formula (S25-2): When voxel ξ is covered by at least one local region, Ω(ξ) = max{F(k,p): ξ∈B(k,p)} (S25-2; When the voxel ξ is not covered by any cell region, Ω(ξ) = 0; Where ξ represents a voxel in the coordinate space of the entire reconstructed volume data; B(k,p) represents the set of voxels corresponding to the local region at position p in the k-th layer, obtained by mapping the forward and reverse coordinates; ξ∈B(k,p) means that voxel ξ belongs to this set of voxels, that is, voxel ξ is covered by the local region at position p in the k-th layer; max{·} means that when the same voxel ξ is covered by multiple local regions, the maximum value of the confidence values ​​F(k,p) of each covered local region is taken; Ω(ξ) is the confidence value at voxel ξ, and the Ω(ξ) of all voxels together constitute the defect confidence field covering the entire reconstructed volume data.

[0117] First, let's explain the design motivation for gating fusion. The reason is that the blind spots of the two types of differences are complementary: cross-layer differences are not sensitive to systemic anomalies that span multiple layers. If an anomaly affects several adjacent layers simultaneously, the difference between layers will not be apparent. Cross-cell differences are not sensitive to process deviations common to the entire package. If cells in the same batch all have the same deviation, the difference between cells will also not be apparent. Real local defects usually destroy both inter-layer consistency and inter-cell consistency simultaneously. Therefore, both must exceed the limit to be judged as high-confidence anomalies. Areas where only a single difference exceeds the limit are more likely to belong to one of the above-mentioned systemic deviations or residual noise. They are downweighted rather than directly eliminated, preserving their opportunity to be merged with neighboring evidence in step S3 three-dimensional aggregation.

[0118] Specifically, continuing the previous example: Let the threshold for determining the inter-layer difference be θL = 0.62 and the threshold for inter-cell difference be θC = 0.58 (the calibration process is shown in Example 2). At position p in the previous example, the aggregated inter-layer difference is 0.71 ≥ θL, and the aggregated inter-cell difference is 0.645 ≥ θC. Both exceed the limits simultaneously, which is determined to be a high-confidence anomaly. A high-confidence value is taken at this position in the defect confidence field. The specific fusion function for the confidence value is preset, for example, the minimum value of the two normalized aggregated differences. If only a single difference exceeds the limit, the confidence value at this position is multiplied by a weighting coefficient λ, which is a preset value, for example, 0.3.

[0119] The results of gating fusion are written back to the three-dimensional space point by point through a forward and reverse coordinate mapping table: the confidence value of each local region on the stacked image sequence is assigned to the voxel it covers by the reverse mapping, and the confidence value of the voxel not covered by any cell region is 0, which finally forms a defect confidence field covering the entire package reconstructed volume data.

[0120] It should be noted that surface unfolding transformation is preferred but not mandatory. In practice, however, surface unfolding transformation can be omitted. Instead, the region correspondence between adjacent electrode layers within the same cell and between different cells can be established directly in the reconstructed data based on the extracted central surface or electrode layer structure orientation. For any local region, the corresponding region on the adjacent electrode layer is found along the normal direction of its central surface to calculate the cross-layer difference. Based on the correspondence between the design positions of each cell in the digital twin reference, the corresponding region of other cells of the same model is found to calculate the cross-cell difference. Propagation aggregation and gated fusion are also performed directly in three-dimensional space along the layer sequence direction of the electrode layers. This alternative method eliminates the unfolding and mapping backfilling steps, but at the cost of establishing and retrieving the correspondence in curved coordinates, making it more cumbersome. It is suitable for cases where the electrode layer curvature is gentle (e.g., large-size stacked cells). The two methods are completely consistent in their comparison-aggregation-fusion processing logic.

[0121] The specific implementation method of step S3 will be described below.

[0122] Step S3 includes the following steps: Step S31: Binarize the defect confidence field according to the preset confidence threshold, perform three-dimensional connected component analysis, aggregate spatially connected anomalous voxels into candidate defects, and merge candidate defects that are adjacent to each other and of the same origin according to the adjacency determination rule.

[0123] Specifically, the defect confidence field is binarized according to equation (S31-1): When Ω(ξ)≥τ, χ(ξ)=1; when Ω(ξ)<τ, χ(ξ)=0 (S31-1); Where χ(ξ) represents the value of voxel ξ after binarization, χ(ξ)=1 indicates that voxel ξ is an abnormal voxel, and χ(ξ)=0 indicates that voxel ξ is a normal voxel; τ is a preset confidence threshold, for example, 0.5; Ω(ξ) is the confidence value of voxel ξ obtained by equation (S25-2).

[0124] Subsequently, a three-dimensional connected component analysis is performed on all anomalous voxels, and spatially interconnected anomalous voxels are aggregated into candidate defects, according to equation (S31-2): {Y1, Y2,...}=N({ξ∣χ(ξ)=1}) (S31-2); Where {ξ∣χ(ξ)=1} represents the set of all anomalous voxels; N(·) represents the three-dimensional connected component analysis performed according to the 26-neighborhood connectivity criterion, which divides all anomalous voxels into several sets of maximal connected voxels; Yμ represents the μ-th candidate defect, where μ is the number of the candidate defect, and each candidate defect is a set of maximal connected voxels.

[0125] Three-dimensional connected component analysis is an existing technology. It is designed to group anomalous voxels together based on whether they are connected in three dimensions. In this way, previously independent anomalous voxels are connected into anomalous bands, which may represent defect locations.

[0126] For candidate defects that are adjacent to each other and of the same origin, they are merged according to the adjacency determination rule of equation (S31-3): min{‖ξ-ξ′‖: ξ∈Yμ, ξ′∈Yμ′}<δ (S31-3); Where Yμ′ represents another candidate defect numbered μ′; ξ′ represents a voxel in Yμ′; ‖ξ-ξ′‖ represents the spatial distance between voxels ξ and ξ′; min{·} represents the minimum value among all voxel pairs of distances between Yμ and Yμ′, i.e., the minimum spatial distance between the two candidate defects; δ is a preset merging distance, for example, 2 to 3 voxels. When equation (S31-3) holds, and Yμ and Yμ′ are located in the same cell region and have the same grayscale shift direction (high similarity), Yμ and Yμ′ are determined to be adjacent and homologous, and the two are merged into one candidate defect.

[0127] Step S32: Determine the three-dimensional location of each candidate defect in the battery pack, the cell region number to which it belongs, and the layer sequence number of the electrode layer in which it is located. Send the defect results to the background and generate an early warning report.

[0128] refer to Figure 5 Specifically, the three-dimensional location is determined by the centroid coordinates of the candidate defect voxel set and its bounding box; the cell region number is directly obtained from the cell region determined in step S2; the layer sequence number of the electrode layer is obtained by looking up the layer sequence number in the stacked image sequence from the voxel of the candidate defect using a forward and reverse coordinate mapping table, and the start and end layer sequence numbers of candidate defects spanning multiple layers are recorded. These three pieces of positioning information allow the re-inspection personnel to directly follow the map and accurately perform the re-scanning or disassembly actions down to the specific layer of the specific cell.

[0129] The early warning report includes the warning location and warning level. The warning location is the three-dimensional location of the candidate defect, the cell region number, and the electrode layer sequence number. The warning level is the risk level of the candidate defect, which is obtained by mapping the volume of the candidate defect according to preset volume levels. For example, a volume smaller than the first threshold level is judged as low risk, a volume between the first and second threshold levels is judged as medium risk, and a volume greater than the second threshold level is judged as high risk. The first and second threshold levels are preset values. The volume is the voxel count of the candidate defect multiplied by the volume of a single voxel. The backend refers to the quality management server or terminal of the production line. The inspection report is a report that includes the above warning location and risk level. It is generated and archived along with the early warning report for re-inspection and quality traceability.

[0130] Thus, Example 1 completes the entire chain from projection acquisition, metal artifact correction and reconstruction, surface unfolding comparison, propagation aggregation and gating fusion to three-dimensional aggregation classification and grading: the digital twin reference body guides the switching of scanning sections during acquisition, provides metal priors during reconstruction, provides cell positioning during comparison, and provides positional semantics during classification and grading.

[0131] Example 2: This example provides the method for constructing the good product reference memory in Example 1, as well as the self-learning calibration and online update method for propagation weights and decision thresholds.

[0132] The good product reference memory provides supplementary reference for step S23. The propagation weight γk and the decision thresholds θL and θC determine the aggregation behavior of equation (1) and the difference boundary of step S25, respectively. The quality of the three directly determines the false alarm and false negative levels of the detection.

[0133] Furthermore, the propagation weight and judgment threshold are updated online as good product data from the production line accumulates: The local structural descriptions of the stacked image sequences of battery packs that were determined to be defect-free and confirmed to be qualified after re-inspection are added to the good product reference memory, and the propagation weights and judgment thresholds are recalibrated accordingly.

[0134] The specific steps are as follows: Step A1: Determine the warehousing conditions for good quality battery packs.

[0135] Battery packs that meet all three of the following conditions can be accepted into the warehouse as good products: (1) The battery pack was determined to be defect-free by the method of Example 1, that is, the overall judgment conclusion is qualified; (2) The battery pack passed the final inspection (electrical performance test, airtightness test, etc.) after leaving the production line; (3) The battery pack is confirmed to be qualified by manual re-inspection according to the preset sampling ratio. The re-inspection method is high-resolution re-scanning or disassembly sampling.

[0136] The above three conditions are indispensable because if the data is entered into the database based solely on the method's own judgment, defective packages that are missed will be treated as good products and thus contaminate the memory database and calibration standards.

[0137] Step A2: Organize the memory bank records using the cell model as the index key.

[0138] Each record contains a local structural description and grayscale distribution characteristics of the stacked image sequence of each cell in the good battery pack (stored according to layer number and in-plane position index), as well as the scanning segment parameters at the time of acquisition. The reason for using the cell model as the index is that cross-cell comparison and supplementary reference are only meaningful between cells of the same model. The number of layers, interlayer spacing and patterns of cells of different models are not comparable.

[0139] Step A3: Update the memory in a rolling manner: New good product records are continuously added, and records that have exceeded the preset retention period are deactivated. When there is a change in cell model or process such as electrode sheet or electrolyte injection, the records under the index key involved in the change are deactivated and re-accumulated. Before the record volume is restored, the proportion of manual re-inspection can be increased as a transitional measure.

[0140] The self-learning calibration of propagation weight γk is implemented according to the following steps: Step T1: Take out a series of stacked images of several good battery packs from the good product reference memory bank, and calculate the cross-layer difference and cross-cell difference of each position in each layer according to steps S21 to S23 of Example 1.

[0141] Since these battery packs have been confirmed to be defect-free, the differences obtained at this point all come from imaging noise and legitimate individual differences, i.e., noise differences in good products.

[0142] Step T2: Statistically analyze the inter-layer attenuation pattern of noise differences along the sequence direction.

[0143] Specifically, for each layer sequence number k, the correlation rk between the noise difference of the kth layer and the noise difference at the corresponding position of the (k-1)th layer is calculated; the value of rk is in the range of [0,1]. The larger rk is, the more correlated the noise difference of the layer sequence is with the neighboring layers, and the more likely it is to come from slowly varying legitimate individual differences rather than mutually independent random noise.

[0144] Step T3: Determine the propagation weight γk according to rk: γk takes a mapping that is negatively correlated with rk.

[0145] For example, γk = γmax × (1 - rk), where γmax is a preset upper limit, for example, 0.5.

[0146] The reason is that if the noise differences between adjacent layers are highly correlated, it indicates that the differences in that layer sequence have systematic components, and the propagation and aggregation should rely less on evidence from adjacent layers (γk should be small) to avoid the systematic components being repeatedly accumulated and amplified in the recursion; if the noise differences between adjacent layers are almost independent, then more evidence from adjacent layers can be relied on (γk should be large) so that the independent noise can be diluted in the compromise layer by layer.

[0147] For example: a certain type of battery cell has a statistical value of r12=0.40 near the 12th layer. If we take γmax=0.5, then γ12=0.5×(1-0.40)=0.30--this is the propagation weight value used in the example of formula (1) in Example 1.

[0148] The adaptive calibration of the decision thresholds θL and θC is carried out according to the following steps: Step T4: For the noise difference of the good product obtained in step T1, propagation polymerization is completed according to step S24 of Example 1 (using γk calibrated in step T3) to obtain the cross-layer difference distribution and cross-cell difference distribution after polymerization on the good battery pack.

[0149] Step T5: Take the (1-α) quantiles of the two distributions as the judgment thresholds according to the preset false alarm control level α.

[0150] For example, if α = 0.5%, then at most about 50 false positives are allowed per 10,000 local regions on a good product. The 99.5th percentile of the aggregated cross-layer difference distribution of a certain type of battery cell is 0.62, and the 99.5th percentile of the aggregated cross-cell difference distribution of good products is 0.58. Therefore, θL = 0.62 and θC = 0.58. This is the threshold value used in step S25 of Example 1. It should be noted that since the gating fusion in step S25 requires both types of differences to exceed the limit simultaneously, the actual joint false positive rate on good products is much lower than the individual level α. This is precisely why gating fusion can maintain a low false positive rate without raising the threshold.

[0151] Step T6: Whenever a new good product record is imported into the memory bank according to steps A1 to A3, steps T1 to T5 are repeated with the updated memory bank to recalibrate γk, θL and θC.

[0152] In this way, the propagation weight and judgment threshold continuously adapt to the current process status as the production line's good product data accumulates, eliminating the need for repeated manual parameter adjustments.

[0153] refer to Figure 6 Example 3: Based on Examples 1 and 2, Example 3 provides a battery pack defect detection system based on industrial CT, including: The scanning device, arranged at the scanning station, includes a radiation source and detector that move around the battery pack along a spiral trajectory, and a positioning fixture that keeps the battery pack stationary during the scanning process. Among them, the X-ray source, detector and its spiral motion mechanism are existing technologies, and their specific structures will not be further described in this application; the positioning fixture locks the position of the battery pack throughout the scanning process to ensure that the relationship between the spatial pose recorded frame by frame and the battery pack coordinate system is constant, which is the premise for the success of step S14 frame-by-frame forward projection.

[0154] Data storage server, which houses the digital twin reference and the good product reference memory; Specifically, the data storage server stores digital twin references by battery pack model, carries a good product reference memory indexed by cell model, and stores the test reports of each battery pack for quality traceability.

[0155] An algorithm server is used to run the battery pack defect detection method based on industrial CT described in the above embodiments.

[0156] Specifically, the algorithm server receives projection data and frame-by-frame recorded spatial poses from the scanning device, retrieves the digital twin reference body and the good product reference memory from the data storage server, executes steps S1 to S3 as described in Example 1, and outputs a detection report; the data entry, calibration, and online update described in Example 2 are executed as background tasks on the algorithm server and do not occupy the online detection cycle.

[0157] The above are merely preferred embodiments of this application and are not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for detecting defects in battery packs based on industrial CT, characterized in that, include: Step S1: Collect projection data of the battery pack; Based on the digital twin reference, the projection data is alternately corrected for metal artifacts and reconstructed between the projection data domain and the image domain to obtain the whole package reconstructed data. A digital twin reference is an ideal state reference that includes the geometric information, material information, and regional distribution information of each component within the battery pack; Step S2: Based on the design location of each cell in the digital twin reference, determine the area of ​​each cell in the whole package reconstruction data; For any local area within any cell region, compare it with the corresponding area of ​​the adjacent electrode layer within the same cell to obtain the cross-layer difference; By comparing any local area with the corresponding area of ​​other cells of the same model in the same battery pack, cross-cell differences can be obtained; The cross-layer and cross-cell differences are propagated and aggregated along the sequence direction of the electrode layers, and the aggregated cross-layer and cross-cell differences are fused to obtain a defect confidence field covering the entire package reconstruction data. Step S3: Perform three-dimensional aggregation of the defect confidence field to obtain candidate defects, determine the three-dimensional location and risk level of each candidate defect, and generate a battery pack inspection report.

2. The battery pack defect detection method based on industrial CT according to claim 1, characterized in that, Step S1 includes the following steps: Step S11: Place the battery pack at the scanning station. The X-ray source and detector move around the battery pack along a spiral trajectory to collect projection data and record the spatial pose of the X-ray source and detector frame by frame. Step S12: Perform initial reconstruction on the projection data to obtain a preliminary reconstructed volume; Step S13: Spatial registration is performed between the digital twin reference body and the preliminary reconstructed body to obtain the spatial transformation relationship from the digital twin reference body to the preliminary reconstructed body; Step S14: Based on the spatial transformation relationship, transform the geometric model marked as metal material in the digital twin reference body to the coordinate space where the initial reconstructed body is located; For the transformed geometric model, forward projection is performed frame by frame according to the projection geometry determined by the spatial pose recorded frame by frame to obtain the ideal projection shape of the metal component. Based on the ideal projection shape, the contaminated sections of the ray passing through the metal component are marked frame by frame in the projection data. The contaminated sections are then repaired to obtain the repaired projection data. Step S15: Reconstruct using the repaired projection data, and iteratively correct the repair results of the contaminated section by alternating between the projection data domain and the image domain.

3. The battery pack defect detection method based on industrial CT according to claim 1, characterized in that, In step S2: Based on the design location of each cell in the digital twin reference, the area of ​​each cell is determined in the whole package reconstruction data; Before calculating the differences between layers and between cells, a surface unfolding transformation is performed on each cell region. The surface unfolding transformation is: extracting the central surface of each electrode layer within the cell region; For each central surface, an in-plane two-dimensional parameterization is established. The entire reconstructed volume data is resampled within a preset thickness range on both sides of the central surface along the normal direction. Each pole piece layer is flattened into a pseudo-planar image. The pseudo-planar images of all pole pieces are arranged in sequence to form a stacked image sequence in the sequence coordinate system. The differences between layers and between cells are calculated on the stacked image sequence, and the fused results are mapped back to the coordinate space of the whole package reconstructed volume data to form the defect confidence field.

4. The battery pack defect detection method based on industrial CT according to claim 3, characterized in that, For wound cells, directional structural reinforcement is first applied to the cell area. Then, starting from the center area of ​​the core, the center position of each electrode layer is traced outward layer by layer along the direction of the electrode layer structure to obtain a continuous central curved surface of each electrode layer.

5. The battery pack defect detection method based on industrial CT according to claim 3, characterized in that, For laminated cells, the center plane of each electrode is fitted layer by layer along the lamination direction to form the center surface.

6. The battery pack defect detection method based on industrial CT according to claim 3, characterized in that, In step S2, the calculation, propagation aggregation, and fusion of cross-layer differences and cross-cell differences include the following steps: Step S21: Select a local region in any layer of the stacked image sequence using a sliding window method, and extract the gray-level distribution features and local structure description of the local region; Step S22: Search for the most structurally similar reference neighborhood near the corresponding positions of several adjacent layers of this layer, compare the differences between the local region and the reference neighborhood in terms of gray-scale distribution features and local structure description, and obtain the cross-layer difference of the local region. Step S23: In the stacked image sequence of other cells of the same model in the same battery pack, the corresponding reference area is indexed according to the layer number and in-plane position. The differences between the local area and the corresponding reference area in gray-scale distribution features and local structure description are compared to obtain the cross-cell difference of the local area. Step S24: Along the layer sequence direction from the inner layer to the outer layer and then from the outer layer to the inner layer, the cross-layer difference and cross-cell difference of each layer are successively merged with the aggregation results of the adjacent aggregated layers according to the propagation weight to complete the propagation aggregation; The propagation weights are obtained by self-learning calibration based on the interlayer attenuation law of noise differences in the stacked image sequence of good battery packs. Step S25: When the cross-layer difference after aggregation and the cross-cell difference after aggregation in any local region both exceed their respective judgment thresholds, it is judged as a high-confidence anomaly; The high-confidence anomaly is mapped back to the coordinate space of the whole package reconstructed volume data to obtain the defect confidence field.

7. The battery pack defect detection method based on industrial CT according to claim 1, characterized in that, Step S3 includes the following steps: Step S31: Binarize the defect confidence field according to the preset confidence threshold, perform three-dimensional connected component analysis, aggregate spatially interconnected anomalous voxels into candidate defects, and merge candidate defects that are adjacent to each other and of the same origin according to the adjacency determination rule. Step S32: Determine the three-dimensional location of each candidate defect in the battery pack, the cell region number to which it belongs, and the layer sequence number of the electrode layer in which it is located. Send the defect results to the background and generate an early warning report.

8. The battery pack defect detection method based on industrial CT according to claim 7, characterized in that, The defect confidence field binarization method is as follows: When Ω(ξ)≥τ, χ(ξ)=1; when Ω(ξ)<τ, χ(ξ)=0; Where Ω(ξ) is the confidence value at voxel ξ, χ(ξ) represents the value of voxel ξ after binarization, χ(ξ)=1 indicates that voxel ξ is an abnormal voxel, χ(ξ)=0 indicates that voxel ξ is a normal voxel; τ is the preset confidence threshold.

9. The battery pack defect detection method based on industrial CT according to claim 8, characterized in that, The merging method for candidate defects is as follows: Perform three-dimensional connected component analysis on all anomalous voxels and aggregate spatially connected anomalous voxels into candidate defects; {Y1, Y2,...}=N({ξ∣χ(ξ)=1}); Where {ξ∣χ(ξ)=1} represents the set of all anomalous voxels; N(·) represents the three-dimensional connected component analysis performed according to the 26-neighborhood connectivity criterion; Yμ represents the μ-th candidate defect, where μ is the number of the candidate defect, and each candidate defect is a maximal connected voxel set; min{‖ξ-ξ′‖: ξ∈Yμ, ξ′∈Yμ′}<δ; Where Yμ′ represents another candidate defect numbered μ′; ξ′ represents a voxel in Yμ′; ‖ξ-ξ′‖ represents the spatial distance between voxels ξ and ξ′; min{·} represents taking the minimum value among all voxel pairs of distances between Yμ and Yμ′, that is, the minimum spatial distance between the two candidate defects; δ is the preset merging distance; when min{‖ξ-ξ′‖:ξ∈Yμ,ξ′∈Yμ′}<δ holds, Yμ and Yμ′ are merged into one candidate defect.

10. A battery pack defect detection system based on industrial CT, characterized in that, include: The scanning device, arranged at the scanning station, includes a radiation source and detector that move around the battery pack along a spiral trajectory, and a positioning fixture that keeps the battery pack stationary during the scanning process. Data storage server, which houses the digital twin reference and the good product reference memory; An algorithm server for running the battery pack defect detection method based on industrial CT as described in any one of claims 1 to 9.

Images