Prismatic Battery Cell Appearance Inspection

This Category positioning

All-in-one appearance inspection machine for square lithium-ion battery cell production lines: multi-surface, multi-defect-type automatic in-line inspection

Prismatic battery cells use several different materials and complex surface structures. Different areas of the cell Color, texture, reflection characteristics, surface topography and defect behavior show clear differences, so a single vision imaging method can hardly cover all inspection requirements.

This system is therefore designed for different inspection zones Dedicated visual inspection module, and combines different imaging methods with AI vision algorithms to achieve multi-dimensional inspection from 2D appearance to 3D topography. The complete machine performs coordinated multi-station inspection according to the actual inspection requirements, carrying out automated visual checks on several key surfaces of the prismatic cell.

inspection Pain Points

Why Prismatic Cells Cannot Be Handled with a Single Imaging Method

Core Technical Solution

A combination of six technologies, not a single imaging method

The system fuses 2.5D visual inspection, 3D visual inspection, photometric stereo imaging, AI defect inspection, machine vision algorithms and vision software platform, achieving automatic inline inspection of multiple types of defects on multiple surfaces of prismatic lithium battery cells.

inspection Scope

Four Core Inspection Areas and Their Main Inspection Scope

Depending on the actual inspection requirements, the complete machine can achieve Multi-station collaborative inspection, performing automated visual inspection on several key surfaces of prismatic battery cells.

inspection Object

Defect Grading and Recognizable Inspection Objects

Addressing quality issues of different grades in prismatic lithium battery cell production, the system can establish, according to the customer's quality standards, Defect classification and grading rules.

The specific defect categories and acceptance criteria can be Product model, customer inspection specifications and actual defect samples are configured and validated.

Inspection Objects Covered

Typical defect

Defect Types, Judgement Thresholds, and AI Class Definitions

Typical Defect Judgement Rules

For different defect types, inspection thresholds can be set according to product quality standards. For example:

Detectable Surface Defect Types

AI Model Class Definitions

through "Segmentation + Classification" two-stage recognition approach, defect regions can be located first and then classified in detail.

inspection Capabilities

Core inspection performance metrics given in the solution

AI vision algorithm

A light-fusion hybrid algorithm network and CNN + attention architecture for high-speed production lines

To meet the inspection requirements for surfaces of different materials on prismatic lithium battery cells, the system uses Custom AI vision algorithm. Algorithm design does not simply rely on a single model; it builds the corresponding visual inspection workflow from "product material → optical imaging → image features → defect type → inspection requirement".

For blue film and similar Large field of view, high resolution inspection areas, so the model inference time must be controlled while guaranteeing inspection accuracy; the system therefore uses an algorithm architecture that fuses local and global features.

1. Algorithm Design for High-Speed Production Lines

• The complete machine achieves a production line cycle time of 40 PPM;
• Blue film inspection has a large data volume, comprising 5 Inspection Stations and must handle multiple high-resolution image streams simultaneously under extreme conditions;
• A single image channel can reach up to 8192 × 4096;
• The total processing time limit is about 1.5 seconds, of which the AI model inference time is allocated about 1 second.

The algorithm therefore needs to solve three problems at the same time: global information (understanding the image structure and defect context over a larger range), local details (effectively extracting local defects such as tiny scratches, bumps and white spots), inference efficiency (controls the model's computation load under limited computing power to meet the real-time inspection needs of high-speed production lines).

2. CNN + Attention Mechanism Hybrid Algorithm Architecture

Traditional CNN networks have strong local feature extraction capability, but a larger receptive field is only obtained gradually as network depth increases; Transformer-type networks can capture global image features earlier, but consume more computing resources on high-resolution industrial images. This system therefore uses CNN fused with attention mechanism hybrid algorithm network.

• Multi-scale feature extraction: CNN is used for multi-scale feature extraction, fusing feature information from different levels.
• Token feature construction: the multi-scale features are further organized into token features and fed into the attention encoding module.
• Global feature understanding: attention modules enhance the model's ability to understand large-scale image structure and defect context.
• Local and global feature fusion: strengthened skip connections fuse local defect features obtained from shallow layers with global features from deep layers.

3. Model Inference Performance Comparison

Test input size: 1 × 3 × 8192 × 4096

Based on project test data, the light-fusion hybrid algorithm network under these input conditions balances High-resolution image processing capability and model inference efficiency.

IV. AI Model Lightweighting

Vision Software Platform

Edge inference management platform, training platform and sample tools

The edge inference management platform is deployed on site inspection equipment and handles Visual inspection workflow, equipment control, algorithm communication, and production data management. The platform supports no-code visual workflow orchestration and packages a range of industrial vision processing capabilities.

Core Software Modules

Core Functions of the Visual Inspection Software

• Product recipe management Supports recipe management for different product models and inspection parameters.
• Batch management Records the products and inspection information corresponding to different production batches.
• Real-time inspection: Displays production images, inspection results and defect information in real time.
• Motion and I/O Control: achieves coordinated control between the visual inspection workflow and automation equipment.
• AI algorithm scheduling: Handles AI model invocation, image transmission and inspection result reception.
• User Permissions: Manages system operation permissions by user role.
• Log tracking: Records equipment operation, user actions and inspection process information.
• Data review: supports queries and playback of production inspection data and related images.

Cloud-based model training platform

In industrial vision projects, Large sample volumes, complex defect types, continuous model iteration and similar requirements, a model training platform can be provided. The platform supports industrial AI algorithm tasks such as image classification, object detection, semantic segmentation and OCR, and builds up industrial defect samples and algorithm models for many types of industrial vision inspection scenarios.

Intelligent industrial defect annotation

For industrial surface inspection scenarios such as lithium battery blue film, graphite film material, mid-frame PVD, and lithium battery posts, intelligent annotation capability can assist in completing Defect sample pre-labeling. For complex samples such as shallow defects and tiny defects, intelligent pre-labeling reduces repetitive manual operations.

AIGC Few-Shot Defect Generation

In industrial vision projects, some defects are inherently low-frequency and the number of real samples is limited. For scarce defects such as shallow scratches on blue film, tiny bumps, white spots and surface anomalies on film, AIGC can be used together with real samples to help expand the training data. The project plan proposes that only About 10 real defect samples As a foundation, it can be used to build a small-sample defect augmentation solution.

Technology Architecture

Multi-Station Collaborative Inspection and Automatic Diverted Unloading

The complete machine uses Multi-station collaborative inspection approach, with the corresponding visual inspection cell configured for each surface structure of the battery cell.

Complete Machine Inspection Workflow

• 01 loading
• 02 2.5D inspection of narrow side faces
• 03 Bottom Surface 2.5D Inspection
• 04 Large Surface 2.5D Inspection
• 05 Top-face 3D and explosion-proof valve inspection
• 06 Large Surface 3D Inspection
• 07 Terminal on-the-fly inspection
• 08 OK / NG judgement
• 09 Automatic diverting unloading

After the material enters the equipment on the loading belt conveyor, it passes through Loading transfer gripper Transfers workpieces, relying on a transfer stage to switch workpieces between different inspection stations.

Automatic diversion after inspection

inspection Method

2.5D Visual Inspection and Photometric Stereo Imaging

2.5D Visual Inspection: Enhancing Micro Surface Defect Recognition

For surface defect inspection requirements such as blue film and aluminum shells, the system uses 2.5D Visual Inspection technology, enhancing faint surface anomalies that conventional 2D images cannot make stand out. Multi-dimensional image information captures the product's surface features and helps identify the following defects:

Photometric stereo imaging technology

The system uses a photometric stereo imaging solution that acquires multi-dimensional image information from the product surface under illumination from different directions; it can output several kinds of image data and also supports Strobe image acquisition mode.

For low-contrast defects such as pits, surface foreign matter and bubbles, it can be combined with Fusion analysis of multi-mode images. By complementing different imaging information, the visual features of some weak defects can be further enhanced, providing richer input data for subsequent AI defect recognition.

Equipment and specifications

Single-station vision hardware configuration reference

Taking a single-station visual inspection module as an example, the configuration in the solution is as follows:

The specific hardware configuration depends on Product dimensions + defect size to detect + field of view + production line speed + imaging requirements are designed and selected.

Applicable Industry

Battery Cell Types and Production Scenarios Mainly Addressed

Can be configured according to different battery cell models and production line layouts Vision module, algorithm and automation structure customization.

Project measurements and Implementation Data

The test and implementation benefit figures given in the solution

Model inference performance comparison (test input 1 × 3 × 8192 × 4096)

Benefits of Model Iteration

Implementation Workflow

Full Inference Flow and Edge-Cloud Closed Loop for AI Defect Inspection

The system uses "Segmentation → Cropping → Classification → Parameter Extraction → Business Grading" the complete AI inspection workflow.

• AI Defect Segmentation: the raw image is fed into the AI defect segmentation module, which segments image regions and defect categories while distinguishing background regions, defect regions and plastic regions.
• Mask processing: generates a Mask from the segmentation result and extracts the bounding rectangle of the defect region.
• Defect region expansion: the defect area is cropped with an outward expansion of about 1.3x; for areas smaller than the target size, image cropping is completed by padding the edges.
• Image Resize Resizes the cropped defect image to the model input size.
• AI Defect Classification: the processed defect area is fed into an AI classification model for further identification of the defect type.
• Defect parameter output: The system outputs pixel-level defect results, including defect length, defect width, defect area, defect center coordinates, defect box position, and defect category.
• Result fusion: final grading combines the defect centroid + region judgement + defect category + size threshold + production line business rules.

AI Vision Edge-Cloud Closed Loop

Through coordination between the equipment side and the model training platform, a complete AI model iteration loop can be formed:

• 01 Sample acquisition
• 02 Sample analysis
• 03 Intelligent annotation
• 04 Model Training
• 05 Model evaluation
• 06 Model encapsulation
• 07 On-site deployment
• 08 Production inspection
• 09 Data feedback
• 10 Continuous model optimization

Why Choose EEK

The core capabilities this solution provides at the product level

Scalable Core Technical Capabilities

This Category Solution

Related solution pages that complement this solution

Related Technical Guides

This Category Document Checklist

Confirmed items and outstanding items

Common Question

Common Questions About Appearance Inspection of Prismatic Lithium Battery Cells

Submit sample testing

Send us the battery cell model, defect samples, inspection requirements and production line cycle time; a solution engineer will determine which category of inspection problem it belongs to and give recommendations on the imaging method and equipment configuration.