Electrode sheet visual inspection method

Abstract

This invention discloses a visual inspection method for electrode sheets. Based on the unique identifier of the electrode sheet, original images are acquired for each of its six sides. For each side, several algorithm units are invoked according to the formula sequence to obtain an extractable image and a list of outline objects. Defect feature calculation and threshold judgment are then performed on the extractable image to obtain the conclusion for that side. The conclusions of the six sides of the electrode sheet with the same identifier are merged into a total judgment, and an overlay display image and structured data are output. The total judgment controls the sorting of electrode sheets. Advantages include: when inspecting products undergoing model changes, only the formula parameters need to be modified without modifying any source code; no modification, recompilation, or redeployment is required; the output of each step is a measurable physical quantity, not an abstract confidence level; the complete data link from defect discovery to qualitative assessment is traceable, allowing process engineers to optimize upstream stamping or plating processes; the merged conclusions of the six sides must belong to the same physical object, rather than a simple summary of discrete image processing results.

Description

Technical Field

[0001] This invention relates to the field of electrode sheet quality inspection technology, specifically to a visual inspection method for electrode sheets. Background Technology

[0002] The appearance quality of cylindrical or similarly structured electrode plates (also known as electrode caps) in new energy batteries, after processes such as stamping, plating, and sealing ring assembly, directly affects the battery's sealing performance and safety. Common defects include: scratches on the plating, dents, dirt, foreign objects, rust and discoloration, damaged or missing sealing rings, poor top surface markings, circumferential burrs and bumps, and geometric deformation. High production line speeds mean that manual visual inspection can lead to missed inspections, inconsistent standards, and untraceable data.

[0003] To avoid the problems associated with manual visual inspection, vision inspection based on industrial cameras has attracted widespread attention. However, existing vision inspection based on industrial cameras suffers from the following problems: Rigid workflow: Inspection logic is strongly coupled with the project, often requiring code modifications for model changes, resulting in high maintenance costs; Poor module reusability: Repetitive development of preprocessing, segmentation, positioning, and image acquisition processes leads to poor consistency; Weak multi-station collaboration: Lack of unified face indexing, time alignment, and result merging; Insufficient interpretability: Defects and physical quantities (area, contrast, position, etc.) are difficult to output in a structured manner, hindering process traceability; These issues urgently require improvement. Summary of the Invention

[0004] One of the technical problems to be solved by this application is to overcome the defects of the above-mentioned related technologies and provide a visual inspection method and equipment for electrode sheets based on unique identifiers and traceable processes with good module reusability and low maintenance cost when the product is changed.

[0005] The technical solution adopted by this electrode sheet visual inspection method to solve the technical problem is as follows: An electrode sheet visual inspection method, comprising the following steps: Based on the unique identifier PartID of the electrode sheet, the original images I_k of its six sides are obtained respectively, k∈{TOP,BOTTOM, SIDE_1, SIDE_2, SIDE_3, SIDE_4}; For each face, several algorithm units are called in the order of Recipe_k to obtain the extracted image R_k and the list of outline objects; Then, perform defect feature calculation and threshold determination on the extracted image R_k to obtain the conclusion Result_k of this surface; The conclusions of the six sides of the same PartID electrode sheet are merged into the total decision Result_total, and the overlay display diagram and structured data are output. Electrode sheet sorting is controlled by the total decision result_total.

[0006] Preferably, the algorithm unit has five components, namely: Size scaling unit: Performs scaling down, enlarging, or custom width and height adjustments on the input original image and outputs a normalized image; Color to Single Channel Unit: Converts the normalized image to grayscale and outputs it in a single channel; Grayscale to Binary Image Unit: Performs fixed thresholding, Otsu's method, trigonometric method, or mean / Gaussian adaptive thresholding on the grayscale image and outputs a binary image B_k; Direct contour detection unit: Extracts contours from binary or grayscale images and generates multiple Object_ITEM, including polygon point columns (pts) and hierarchical relationships; Shape Extraction Image Unit: Based on the object outline, the target region is cropped from the original image, normalized image, or grayscale image, and the extracted image R_k is output.

[0007] Preferably, the five algorithm units inherit from the common base class Basic_ACT, and during the registration phase, they write a globally unique UnitCode, display name Name, description Desc, algorithm unit number KNo, and input parameter mode FunRun_InPara. When the five algorithm units run, the scheduler calls the virtual interface FunRun in the order of recipe_k. The input and output are carried in the VALUEX_ITEM and Object_ITEM structures: the image matrix, the polygon point column pts of the object contour, and the key-value measurement field are passed between algorithm units through TAG references.

[0008] Preferably, the algorithm units are passed between each other via TAG references, specifically through the following mechanism: (1) "last" flag: The output of the previous algorithm unit automatically becomes the default input of the next algorithm unit, referencing the list of images or objects written to the Unit structure after the most recent FunRun execution; (2) "Uxx" flag: refers to the output of the specified algorithm unit number KNo, allowing skip-level references; (3) "Object:last" tag: refers to the set of output contours of the algorithm unit that last generated the list of objects.

[0009] Preferably, the recipe_k for each face is a JSON array, with each element containing: algorithm unit number KNo and input parameter coverage values.

[0010] As a preferred approach, the conclusions of the six sides of the same PartID are merged into a total decision Result_total, which specifically includes: Step M1, PartID Association and Face Cache: Using the unique identifier PartID as the unique key, a six-face result cache table for the electrode sheet is established in memory. After each face is tested, Result_k is written to the corresponding face slot. Step M2, Integrity Verification: Before merging, first verify whether the conclusions of all six sides are complete; Step M3, Timing Alignment: Confirm that the conclusions of the six sides belong to the same PartID within the time window Δt; Step M4, Overall Acceptance Judgment - Dual Mode Configurable: Provides two judgment modes, which can be selected by process configuration as either the overall judgment mode or the weighted scoring judgment mode; Step M5, NG Root Cause Locator and Structured Output: When the overall pass / fail judgment is NG, the NG root cause is automatically marked, forming a structured traceability record; Step M6, Data Persistence and External Communication: Write the total judgment result_total, six-sided result_k, measurement values ​​of each face, and defect overlay map to the local database or upload them to the MES database; output sorting signals to the PLC and record a complete inspection log.

[0011] Preferably, the all-and-match determination mode is: OK_total = ∧ (k∈K) OK_k; Where k = {TOP, BOTTOM, SIDE_1, SIDE_2, SIDE_3, SIDE_4}; means that the pass is only allowed if all six sides are OK, and if any side is NG, it is always judged as NG. OK means pass, and NG means fail. The weighted scoring determination mode is as follows: S = Σ (k∈K) w_k·s_k, OK_total = (S ≥ S_min); Where: w_k is the face weight coefficient, reflecting the relative importance of that face to product safety; s_k is the face score; S_min is the pass threshold, which is reverse-calibrated by the yield target.

[0012] Compared with related technologies, this electrode visual inspection method has the following advantages: First, when inspecting a product model change, only the formula parameters need to be modified without modifying any source code, and there is no need to modify, recompile, or redeploy; Second, the output of each step is a measurable physical quantity, rather than an abstract confidence level, and the complete data link from defect discovery to qualitative analysis is traceable; Third, process engineers can use this to optimize upstream stamping or plating processes; Fourth, the conclusions from the six sides must be attributed to the same physical object, rather than a simple summary of discrete image processing results. Attached Figure Description

[0013] Figure 1 This is a flowchart of the single-sided testing process for this application.

[0014] Figure 2 This is a flowchart summarizing the conclusions of the six aspects of this application. Detailed Implementation

[0015] First, those skilled in the art should understand that these embodiments are merely used to explain the technical principles of the embodiments of this application and are not intended to limit the scope of protection of the embodiments of this application. Those skilled in the art can make adjustments as needed to adapt to specific application scenarios.

[0016] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0017] This invention designs a visual inspection method for electrode sheets, applied to industrial machine vision inspection and quality control in intelligent manufacturing processes. Under multi-station or multi-view imaging conditions, the method utilizes computer software to perform image acquisition, geometric and grayscale processing, target contour extraction, region of interest extraction, and appearance defect determination on the six outer surfaces of new energy battery electrode sheets (electrode caps). It also includes result interaction and traceability data output with PLCs, MES systems, or local databases. The flowchart is shown below. Figure 1 and Figure 2 As shown, it includes the following steps: Based on the unique identifier PartID of the electrode sheet, the industrial camera at each detection station acquires the original images I_k of its six sides according to the cycle, k∈{TOP, BOTTOM, SIDE_1, SIDE_2, SIDE_3, SIDE_4}; For each face, several algorithm units are called in the order of Recipe_k to obtain the extracted image R_k and the list of outline objects; Then, perform defect feature calculation and threshold determination on the extracted image R_k to obtain the conclusion Result_k of this surface; The conclusions of the six sides of the same PartID electrode sheet are merged into the total judgment Result_total, and the superimposed display diagram (also known as the defect superimposed diagram) and structured data are output. Electrode sheet sorting is controlled by the total decision result_total. The algorithm unit consists of five parts, namely: Size scaling unit: Performs scaling down, enlarging, or custom width and height adjustments on the input original image and outputs a normalized image; KNo: 1100; UnitCode example: SIZE_ZOOM_V20220604; Function: Performs scaling down, scaling up, or custom width and height operations on the input raw image I_k (RGB or GRAY), allowing subsequent thresholding and contour parameters to be shared across multiple stations and resolutions; Main input parameters: Img (image to be processed), Mode (zoom out / zoom in / custom size), ZoomRatio or DiyW, DiyH (magnification or pixel-level target size); Output: The normalized image, used by algorithm unit 1105; Physical meaning: Mapping the differences in object distance and pixel size at different workstations to a unified pixel scale, making the same set of segmentation thresholds comparable.

[0018] Color to Single Channel Unit: Converts the normalized image or the original image into a grayscale image and outputs it in a single channel; KNo: 1105; UnitCode example: RGB2GRAY_V180806; Functions: Convert RGB to grayscale, or extract H / S / V (hue / saturation / brightness), specify color channels, color cast enhancement channels, etc., to highlight coating reflections, color differences, or stains; Key input parameters: Img (RGB or grayscale input), TarArea (overall area or within the object outline), Mode (GRAY, H, S, V, saturated / unsaturated, etc.); Output: Single-channel grayscale image, for use by algorithm unit 1110; Physical significance: Compressing multispectral sampling into a scalar field most relevant to defect contrast facilitates subsequent binarization.

[0019] Grayscale to Binary Image Unit: Performs fixed thresholding, Otsu's method, trigonometric method, or mean / Gaussian adaptive thresholding on the grayscale image and outputs a binary image B_k; KNo: 1110; UnitCode example: GRAY_2_BW_V180806; Function: Apply fixed thresholding (thval), Otsu's method, triangulation, adaptive thresholding (mean) / adaptive Gaussian thresholding (adaptiveg) to a grayscale image to obtain a binary image B_k; Main input parameters: Img (grayscale image), Mode (thval / ostu / triangle / adaptive / adaptiveg, etc.); in adaptive mode, block size, threshold offset, etc. Mathematical Example (Adaptive Threshold): For pixel (x,y), let the mean gray level μ_W(x,y) and standard deviation σ_W(x,y) within the neighborhood W, and the contrast coefficient C, then: T(x,y) = μ_W(x,y) - C·σ_W(x,y) B_k(x,y) = 255, if I_gray(x,y) ≥ T(x,y); otherwise, 0; C and block size are determined by the process specifications, and their physical meaning is: to extract local contrast anomalies (such as scratches and dirt) on a background with slowly changing brightness. Output: Binary image, used by algorithm unit 2111 to extract contours.

[0020] Direct contour detection unit: Extracts contours from binary or grayscale images and generates multiple Object_ITEM, including polygon point columns (pts) and hierarchical relationships; KNo: 2111; UnitCode example: CONTOUR_DETECT_DIRECT_200727; Function: Extracts contours from binary or grayscale images, generating multiple Object_ITEMs, including polygon point columns (pts), hierarchical relationships, etc.; ContentImg can be specified for analyzing grayscale features within the contours. Main input parameters: Img (BW / GRAY for edge finding), ContentImg (image for content analysis), FromSideMode (outside / inside of the highlighted boundary), approximation mode, minimum area and other filtering conditions; Output: A list of outline objects, which serves as the Object input to algorithm unit 2202; Physical meaning: Discretize the geometry of electrode edges, steps, sealing ring boundaries, etc. from the pixel array into a measurable set of closed curves.

[0021] Extract Image by Shape Unit: Based on the object outline, crop the target region from the original image, normalized image, or grayscale image, and output the extracted image R_k; KNo: 2202; UnitCode example: ROI_EXTRACT_FROM_IMG_V191212; Function: Based on the outline of the upstream Object, crop the target region from RawImg, with options for margin expansion, rotation / mirror correction, background fill, etc., and output the extracted image (sub-image) R_k; Main input parameters: Object (the outline object output by 2111), RawImg (a color or grayscale image that retains its original resolution), ExtractMode, margins, fill color, etc. Output: Subgraph R_k and its associated geometric information, which will be used by subsequent defect algorithms to perform calculations only within the ROI; Physical meaning: Focusing the "whole image" onto the pixel set of the "effective surface of the electrode sheet" makes even tiny defects appear large enough at local resolution.

[0022] For a fixed face k, the computer program executes according to a strict timing sequence: 1100: I_k → I_k' (Scale / Geometric Normalization) 1105: I_k' → G_k (single-channel grayscale or feature channel) 1110: G_k → B_k (Binary partitioning) 2111: B_k (and optional ContentImg) → Contour set {O_i} 2202: {O_i} + original I_k or I_k' → R_k (extracted graph) Then, the inspection logic is executed on R_k, such as the connected region area threshold, the comparison between the average gray level and the standard deviation window, and the peak detection along the circumferential expansion curve, to obtain OK_k and the defect code.

[0023] The five algorithm units inherit from the common base class Basic_ACT. During the registration phase, they write a globally unique UnitCode, display name Name, description Desc, algorithm unit number KNo, and input parameter mode FunRun_InPara. When the five algorithm units run, the scheduler calls the virtual interface FunRun in the order of recipe_k. Input and output are carried in the VALUEX_ITEM and Object_ITEM structures: image matrix, polygon point column pts of object contour, and key-value measurement fields are passed between algorithm units through TAG references.

[0024] The algorithm units are passed between each other via TAG references, and the specific mechanism is as follows: (1) "last" flag: The output of the previous algorithm unit automatically becomes the default input of the next algorithm unit, referencing the list of images or objects written to the Unit structure after the most recent FunRun execution; (2) "Uxx" marker: refers to the output of the specified algorithm unit number KNo, allowing skipping levels of reference; for example, "U1105" means referencing the image after the 1100 algorithm unit has been executed; (3) "Object:last" tag: refers to the set of output contours of the algorithm unit that last generated the list of objects.

[0025] The recipe Recipe_k for each face is a JSON array, and each element contains: algorithm unit number KNo and input parameter coverage values.

[0026] The conclusions of the six aspects of the same PartID are merged into the total decision Result_total, which specifically includes: Step M1, PartID association and face caching Using PartID as the unique key, a six-sided result cache table for the electrode sheet is built in memory. After each side is tested, Result_k is written to the corresponding side slot. Cache table structure: Physical meaning: PartID corresponds to a physical electrode cap on the production line. The conclusions of the six sides must belong to the same physical object, rather than a simple summary of discrete image processing results.

[0027] Step M2, Integrity Verification Before merging, first verify whether all six conclusions are complete; the specific logic is as follows: (1) Traverse FaceSlot and check whether all six face slots, including TOP, BOTTOM, SIDE_1~SIDE_4, have a valid Result_k; (2) If there is an empty slot (no result for that slot), mark the Status as "missing slot" and execute according to the process strategy: a. Strict mode: Missing surfaces are immediately judged as NG to prevent false acceptance due to missed inspections (applicable to critical surfaces such as sealing surfaces); b. Forgiving mode: For non-critical surfaces (such as SIDE_3, SIDE_4) with missing surfaces, a warning is recorded but no NG is given; for critical surfaces (TOP, BOTTOM) with missing surfaces, no NG is given. (3) The technical effect of integrity verification is to ensure that the test results of the same electrode cover all six outer surfaces and there is no "passing without inspection" caused by blind spots in the test.

[0028] Step M3, Timing Alignment Since there may be differences in imaging speed or transmission delays among the six workstations, it is necessary to ensure that the conclusions from all six surfaces belong to the same PartID within the time window Δt; specifically: (1) Let Δt be the maximum inter-plane delay allowed by the process (e.g., 2 seconds), and take the timestamp of the first arriving Result_k as the reference T_0; (2) Check whether the remaining five Timestamp_k paths satisfy |Timestamp_k - T_0| ≤ Δt; (3) If a certain path times out, the data of that face is determined to be "expired" and is treated as a missing face (rollback to M2); (4) If a PartID switch occurs (adjacent workpieces), the face of the previous PartID that is not complete will be treated as timeout; Physical meaning: On the production line, the electrode sheets flow through six stations at a certain rhythm. The timing alignment prevents the surface result of the previous cap from being mixed with the surface result of the next electrode sheet into the same group (i.e., "serialized parts" or "mixed parts").

[0029] Step M4, Overall Pass / Fail Judgment – ​​Dual Mode Configurable Two judgment modes are provided, which can be selected by the process configuration as either the all-in judgment mode or the weighted scoring judgment mode. The total and decision mode is as follows: OK_total = ∧ (k∈K) OK_k; This means that the system is only approved if all six sides are OK, and the overall decision is OK. If any side is NG, the overall decision is NG. This is suitable for high-security scenarios that require zero defects. Here, OK means pass, and NG means fail, and k = {TOP, BOTTOM, SIDE_1, SIDE_2, SIDE_3, SIDE_4}. The weighted scoring determination mode is as follows: S = Σ (k∈K) w_k·s_k,OK_total = (S ≥ S_min) This indicates that if the weighted score is greater than the pass threshold, the overall judgment is OK; otherwise, the overall judgment is NG. Where: w_k is the surface weight coefficient, reflecting the relative importance of that surface to product safety. For example: TOP surface (sealing surface) w_TOP = 0.3, BOTTOM surface w_BOTTOM = 0.3, and each of the four sides w_SIDE = 0.1; the weight values ​​are determined by process engineers based on the degree of impact of defects on sealing performance. s_k is the face score, which is calculated by combining the number of defects, area, and severity; for example: s_k = max(0, 1 -Σ(defect_area_i / threshold_i)), that is, the smaller the total defect area, the higher the score; S_min is the pass threshold, which is determined by the reverse calibration of the yield target; The advantages of the weighted scoring judgment mode are: it allows minor appearance defects on non-critical surfaces (such as the side) that do not affect the sealing performance, while ensuring that critical surfaces (top and bottom) strictly meet the standards; and it solves the engineering problem of "balancing yield and quality" on the production line.

[0030] Step M5, NG Root Cause Localization and Structured Output When the overall pass / fail rating is NG, the root cause of NG is automatically marked, and a structured traceability record is formed. NG Root Cause Record Table: The aforementioned structured records are simultaneously output to the MES database for quality traceability and SPC statistics, and to the PLC for driving the sorting mechanism to remove the electrode sheet into the NG channel.

[0031] Step M6, Data Persistence and External Communication: Execute after the merge is complete: (1) Write Result_total, six-sided Result_k, measurement values ​​of each face, and defect overlay diagram (NG outline and annotation overlaid on the original diagram) into the local database or upload it to MES; (2) Output sorting signals (OK → release, NG → rejection) to the PLC. The signal type and timing should be consistent with the PLC protocol of the production line. (3) Record complete test logs (including PartID, processing time for each facet, and snapshot of unit parameters) to support post-audit and process optimization.

[0032] The following is a comparison of the formulations for two different electrode cap models within the same software framework, demonstrating that switching models only requires modifying the formulation parameters without altering any source code: Model A: Φ18mm standard cap (six sides fully inspected) Model B: Φ25mm large cap (TOP+BOTTOM only) The only difference between the testing methods for the two types of electrode caps mentioned above is: (1) The parameter values ​​in Recipe_k are different (such as scaling factor, threshold, block size), and are configured by JSON; (2) Some noodle recipes can be empty (such as SIDE_1~4 of model B), and will be automatically skipped during merging; (3) The binary code of the five algorithm units, the Basic_ACT base class, and the TAG reference passing mechanism are completely reused and do not require modification, recompilation or redeployment.

[0033] In other words, for product model changes, this invention only requires process engineers to adjust the formula parameters, with a cycle time measured in minutes. This is a significant technical advantage of the "formulation framework" of this invention compared to the existing "deep learning network" which requires changes to the software coding.

[0034] Verification example: Taking the detection of a "scratch" defect on the SIDE_2 surface of a specific workpiece (PartID=#20260420-00123, Model A standard cap) as an example, the complete data flow from the original image to the merged output is fully demonstrated: The above verification examples show that: (1) The output of each step is a measurable physical quantity (pixel resolution, gray value, contour coordinates, area mm²), rather than an abstract confidence level.

[0035] (2) The complete data link from defect discovery to qualitative analysis is traceable: scratch → pixel area → outline O_2 → area 0.8mm² → exceeding threshold 60% → NG judgment → PLC rejection.

[0036] (3) The existing technology uses YOLOv8 to output "defect category + confidence level", which cannot provide the above-mentioned pixel-level and physical quantity-level data links, nor can it tell engineers "how much the defect area is and how much it exceeds the threshold" - which directly determines whether process engineers can optimize the upstream stamping or coating process accordingly.

[0037] An electrode sheet visual inspection device includes a controller that operates the aforementioned electrode sheet visual inspection method, and an electrode sheet feeding mechanism, an electrode sheet transfer mechanism, a front visual inspection mechanism, a four-sided visual inspection mechanism, a back visual inspection mechanism, an electrode sheet defective sorting mechanism, an electrode sheet tray transport mechanism, and a qualified electrode sheet unloading mechanism, all electrically connected to the controller. The electrode sheet transfer mechanism operates based on a metronome. The four-sided visual inspection mechanism is used to simultaneously photograph the four sides of the electrode sheet in one cycle and identify the side numbers SIDE_1 to SIDE_4 respectively. The controller uses a PLC, and the electrode sheet defective sorting mechanism is controlled and driven by the PLC to reject electrode sheets that are judged as NG to the NG material channel.

[0038] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method of vision inspection of an electrode tab, the method comprising: capturing an image of the electrode tab; and determining a tab width of the electrode tab based on the captured image. Includes the following steps: Based on the unique identifier PartID of the electrode sheet, the original images I_k of its six sides are obtained respectively, k∈{TOP, BOTTOM,SIDE_1, SIDE_2, SIDE_3, SIDE_4}; For each face, several algorithm units are called in the order of Recipe_k to obtain the extracted image R_k and the list of outline objects; Then, perform defect feature calculation and threshold determination on the extracted image R_k to obtain the conclusion Result_k of this surface; The conclusions of the six sides of the same PartID electrode sheet are merged into the total decision Result_total, and the overlay display diagram and structured data are output. Electrode sheet sorting is controlled by the total decision result_total.

2. The electrode sheet vision inspection method according to claim 1, characterized by, The algorithm consists of five units, namely: Size scaling unit: Performs scaling down, enlarging, or custom width and height adjustments on the input original image, and outputs a normalized image; Color to Single Channel Unit: Converts the normalized image or the original image into a grayscale image and outputs it in a single channel; Grayscale to Binary Image Unit: Performs fixed thresholding, Otsu's method, trigonometric method, or mean / Gaussian adaptive thresholding on the grayscale image and outputs a binary image B_k; Direct contour detection unit: Extracts contours from binary or grayscale images and generates multiple Object_ITEM, including polygon point columns (pts) and hierarchical relationships; Shape Extraction Image Unit: Based on the object outline, the target region is cropped from the original image, normalized image, or grayscale image, and the extracted image R_k is output.

3. The electrode sheet visual inspection method according to claim 2, characterized in that, The five algorithm units inherit from the common base class Basic_ACT. During the registration phase, they write a globally unique UnitCode, display name Name, description Desc, algorithm unit number KNo, and input parameter mode FunRun_InPara. When the five algorithm units run, the scheduler calls the virtual interface FunRun in the order of recipe_k. Input and output are carried in the VALUEX_ITEM and Object_ITEM structures: image matrix, polygon point column pts of object contour, and key-value measurement fields are passed between algorithm units through TAG references.

4. The electrode sheet visual inspection method according to claim 3, characterized in that, The algorithm units are passed between each other via TAG references, and the specific mechanism is as follows: (1) "last" flag: The output of the previous algorithm unit automatically becomes the default input of the next algorithm unit, referencing the list of images or objects written to the Unit structure after the most recent FunRun execution; (2) "Uxx" flag: refers to the output of the specified algorithm unit number KNo, allowing skip-level references; (3) "Object:last" tag: refers to the set of output contours of the algorithm unit that most recently generated the list of objects.

5. The electrode sheet visual inspection method according to claim 1, characterized in that, The recipe Recipe_k for each face is a JSON array, and each element contains: algorithm unit number KNo and input parameter coverage values.

6. The method for visual inspection of electrode sheets according to any one of claims 1 to 5, characterized in that, The conclusions of the six aspects of the same PartID are merged into the total decision Result_total, which specifically includes: Step M1, PartID Association and Face Cache: Using the unique identifier PartID as the unique key, a six-face result cache table for the electrode sheet is established in memory. After each face is tested, Result_k is written to the corresponding face slot. Step M2, Integrity Verification: Before merging, first verify whether the conclusions of all six sides are complete; Step M3, Timing Alignment: Confirm that the conclusions of the six sides belong to the same PartID within the time window Δt; Step M4, Overall Acceptance Judgment - Dual Mode Configurable: Provides two judgment modes, which can be selected by process configuration as either the overall judgment mode or the weighted scoring judgment mode; Step M5, NG Root Cause Locator and Structured Output: When the overall pass / fail judgment is NG, the NG root cause is automatically marked, forming a structured traceability record; Step M6, Data Persistence and External Communication: Write the total judgment result_total, six-sided result_k, measurement values ​​of each face, and defect overlay map to the local database or upload them to the MES database; output sorting signals to the PLC and record a complete inspection log.

7. The electrode sheet visual inspection method according to claim 6, characterized in that, The total and decision mode is as follows: OK_total = ∧ (k∈K) OK_k; Where k = {TOP, BOTTOM, SIDE_1, SIDE_2, SIDE_3, SIDE_4}; means that the pass is only allowed if all six sides are OK, and if any side is NG, it is always judged as NG. OK means pass, and NG means fail. The weighted scoring determination mode is as follows: S = Σ (k∈K) w_k·s_k, OK_total = (S ≥ S_min); Where: w_k is the face weight coefficient, reflecting the relative importance of that face to product safety; s_k is the face score; S_min is the pass threshold, which is reverse-calibrated by the yield target.

Images