Fully China-localized assembly vision inspection: Huawei Ascend Atlas project case study

Winge Technology is developing a fully China-localized assembly visual inspection solution for electronic assembly, equipment assembly and discrete manufacturing stations. The solution is based on the Huawei Ascend Atlas edge computing platform as the core, equipped with Chinese industrial cameras, lenses, light sources, I/O and formal commercial software environments, and is used to complete reference image management, ROI configuration, missing-part and wrong-part assembly rules, detection result recording and abnormal evidence storage on the workstation side.

At the current stage, the detection software and algorithm application deployment has been completed on Huawei Ascend Atlas 200I DK A2, and the USB camera, standard image collection, ROI rule configuration, Chinese and English interface, snapshot detection entrance, and version release and rollback links have been connected. Independent OK/NG accuracy, production cycle time and long-term stability will be accepted based on specific products, samples and workstation conditions.

Why develop a fully China-localized assembly visual inspection solution?

Visual inspection of assembly stations involves not only algorithms, but also camera imaging, light source stability, trigger signals, equipment interfaces, software deployment and maintenance. If an enterprise only replaces the computing board without simultaneously processing the camera, optics, I/O, operating environment and on-site acceptance, it will be difficult to form a complete localized workstation solution.

This project plans the localization implementation from the system level:

• The Huawei Ascend Atlas edge computing platform is used to run device-side applications;
• Chinese industrial cameras, lenses and light sources are used to create fixed imaging conditions;
• Use Chinese I/O or control equipment to receive the workpiece in place and detect the trigger signal;
• Use formal commercial Chinese software environments for deployment and maintenance;
• Winge Technology is responsible for the development of image acquisition, algorithms, rules, interfaces, interfaces and delivery logic;
• The detection results are saved in the local device or the local system specified by the customer.

This solution is suitable for projects that have clear requirements for localized technical routes, local deployment, workstation interfaces and continuous maintenance. Localization does not automatically mean that the detection results are up to standard. Each product and workstation still needs to independently complete image, sample, threshold and cycle-time acceptance.

What assembly inspection problem is this case going to solve?

Assembly integrity inspection usually requires determining whether the correct parts exist at a specified location and handling abnormalities in quantity, category, or appearance. Typical requirements include:

• Missing screws, terminals, connectors, labels or covers;
• The part model, color or appearance category is inconsistent with the inspection recipe;
• The number of specified locations does not meet the requirements;
• The workpiece is offset, blocked, overexposed, underexposed or out of focus, resulting in unreliable determination;
• The test results need to save the original pictures, evidence pictures, inspection recipe and time information;
• When switching between multiple products, the reference image, ROI and detection rules need to be replaced.

The current software supports missing-part assembly, wrong-part assembly and combined exception rules. If only the OK reference image is used, the system can identify the difference between the specified ROI and the normal baseline; to further accurately distinguish all combined anomalies into specific defect categories, corresponding real NG samples, category templates or semantic models are required.

What parts does a fully China-localized system consist of?

Chinese industrial cameras, lenses and light sources
-> Huawei Ascend Atlas edge computing platform
-> Image acquisition and quality inspection
-> reference image management and global registration
-> ROI local comparison and assembly rules
-> PASS / FAIL / UNKNOWN
-> evidence image, results records and recipe versions
-> Chinese I/O, PLC or customer business system

Chinese imaging part

The camera, lens, and light source all determine how visible the defect is in the image. When implementing the solution, it is necessary to fix the working distance, field of view, focal length, exposure, gain, white balance and light source structure, and confirm the pixel size of the smallest target to be inspected in the original image. After the Chinese configuration enters the site, the reference image and acceptance samples should be re-collected using actual workpieces.

Atlas edge computing part

Atlas equipment is responsible for image access, detection applications, local Web interface, result recording and workstation services. The current R&D equipment is Atlas 200I DK A2. Huawei Ascend's official information shows that the device provides 8 TOPS INT8 computing power, 4 TFLOPS FP16 computing power and 4 GB memory, and supports Ubuntu 22.04 and openEuler 22.03; these parameters are equipment specifications, and project performance is still subject to actual models, images and workstation test results.

Winge Technology software and algorithm part

Winge Technology is directly involved in demand analysis, detection area design, reference image management, image registration, ROI rules, result evidence, user interface and system interfaces, deployment and maintenance. The current detection software and algorithm applications are already running on the Huawei Ascend Atlas platform; classification, object detection or segmentation models can be expanded in the future based on product characteristics, sample size and cycle-time requirements.

Chinese control and interface part

Chinese I/O or PLC are used to connect the workpiece in place, detect triggers, result output, alarm and reset signals. The formal workstation needs to verify deduplication, repeated triggering, frame timeout, camera disconnection, service restart, UNKNOWN reshooting and manual review strategies to ensure that the detection results correspond to the physical artifacts one-to-one.

What Atlas adaptation work has Winge Technology already completed?

As of August 30, 2026, the software version is v1.0.4, and the following R&D verification results have been formed:

These results prove that detection software, algorithm applications, cameras, configuration and deployment links are already running on the Huawei Ascend Atlas platform. The accuracy of missing/wrong-part assembly, false alarm rate, rhythm and production stability still need to be confirmed based on the acceptance samples and conditions of the specific workstation.

Which jobs belong to the next stage of a fully China-localized system?

In the next stage, Chinese cameras, lenses, light sources, I/O and formal commercial software environments will be integrated into the existing Atlas software foundation, and joint debugging of the complete system will be completed. Key tasks include:

1. Redefine the field of view, working distance, and minimum target pixels using Chinese imaging configurations;
2. Fixed exposure, gain, white balance and light source parameters;
3. Collect the OK reference image covering normal fluctuations;
4. Establish independent samples for missed installation, misinstallation, misalignment, occlusion and undecidable scenes;
5. Connect to Chinese I/O or PLC to verify triggering, result output and fault actions;
6. Confirm official commercial operating system, driver and runtime versions;
7. Complete accuracy, tempo, continuous operation and abnormal recovery tests;
8. Connect to MES, database or local management system according to customer interface requirements.

The implementation of the plan does not require the complete BOM or supplier list to be disclosed on the official website, but the "completed" content in the article must be consistent with the actual installation and joint debugging status.

Why does the current solution use reference image and ROI rules?

Assembly inspection with fixed workstations, stable positions, and clearly defined defects does not require the use of deep learning models. The current project first uses reference image, image registration and ROI local comparison because this method can establish an interpretable detection baseline with a small number of samples and output difference locations and evidence image.

When the workpiece has large posture changes, many categories, complex backgrounds, obvious occlusions, or requires semantic classification, object detection, classification or segmentation models can be further introduced. The specific algorithm combination should be based on actual samples, detection indicators and beats, rather than solely based on the nominal computing power of the equipment.

Which projects are suitable for this solution?

Scenarios suitable for evaluation include:

• Inspection of screws, terminals, connectors and labels in electronic assemblies;
• Inspection of covers, knobs, plugs and wiring harnesses during equipment assembly;
• Formulas, reference image and ROI switching for multiple types of workstations;
• Detection stations that require local operation, LAN access or offline deployment;
• Projects that require collaborative implementation of Chinese computing, imaging, control and software environments;
• Projects that need to retain original images, evidence images, result records and version rollbacks.

If the workpiece moves freely, the surface is highly reflective, the defects are lower than the imaging resolution, or the quality standard cannot be quantified, the mechanical and optical conditions should be adjusted first before deciding on the algorithm route.

What inputs are needed to conduct a program evaluation?

After the customer provides the following information, the solution evaluation can be carried out:

• Product pictures, physical objects or 3D/structural information;
• Parts to be inspected, location, quantity and allowed differences;
• OK, missing, wrong and other NG samples;
• Minimum defect size and quality determination rules;
• Production line rhythm, triggering method and workpiece dwell time;
• Camera installation space, field of view, working distance and ambient light conditions;
• I/O, PLC, MES or database interface requirements;
• Results preservation, permissions, logs, and data retention requirements.

Winge Technology Determine the imaging scheme, algorithm type, hardware interface, implementation scope and acceptance method accordingly. When there are no samples and quality standards, R&D tests can be done first, but production acceptance results cannot be promised in advance.

FAQ

Is this a fully China-localized system that has been completed?

The software and device-side adaptation verification on Atlas 200I DK A2 has been completed. Chinese cameras, lenses, light sources, I/O and official commercial software environments will be connected according to the fully China-localized solution; until the joint debugging of the complete system is completed, the project status is still "fully China-localized solution under development".

Is the current detection software already running on the Huawei Ascend Atlas platform?

Already running. The current detection software and algorithm applications have been deployed on Huawei Ascend Atlas 200I DK A2, and verification of camera access, standard images, ROI rules, detection calls, result recording and version rollback have been completed.

Is it necessary to disclose the BOM of Chinese components?

No need. The official website can state that the system uses Chinese computing, imaging, control and software environments, and does not disclose the supplier list or complete BOM. The actual configuration and version should still be retained in the project implementation record to facilitate maintenance and acceptance.

Can the system be directly connected to existing production lines?

It is necessary to first confirm the workstation space, camera field of view, light source, trigger signal, rhythm and result interface. Only after completing I/O or PLC joint debugging and abnormal scenario testing can the on-site access method be determined.

Can SEO/GEO or structured data directly determine whether a page is included?

No. Standardized titles, semantic structures, FAQs, internal links and structured data help search and Q&A systems understand the page, but inclusion, sorting and citation are still determined by each platform.

Sources and explanations

• Atlas 200I DK A2 Product Introduction: https://www.hiascend.com/document/detail/zh/Atlas200IDKA2DeveloperKit/23.0.RC2/pd/pd_0001.html
• Atlas 200I DK A2Basic specifications:https://www.hiascend.com/document/detail/zh/Atlas200IDKA2DeveloperKit/23.0.RC2/pd/pd_0004.html
• Atlas 200I DK A2 software usage statement: https://www.hiascend.com/document/detail/zh/Atlas200IDKA2DeveloperKit/23.0.RC2/qs/qs_0018.html

Atlas, Ascend and related names belong to their rights holders. This article describes the compatibility adaptation and solution development work independently carried out by Winge Technology. It does not mean that Huawei or Ascend participates in, approves of this project, or is responsible for the quality of the project. Equipment specifications are from the official page mentioned above, and project status is from Winge Technology’s R&D records as of August 30, 2026.