From Atlas developer-kit validation to a China-localized production vision system

To move from algorithm and software verification on the Atlas developer kit to a fully China-localized mass production solution, it is necessary to replace or confirm the official commercial hardware and software environment, access Chinese cameras, lenses, light sources and I/O, and complete real sample, rhythm, stability, safety and on-site interface acceptance. Being able to run on the development board can only prove that the technical route has the basis for continued implementation, and cannot replace commercialization and production release.

Winge Technology's current assembly visual inspection software and algorithm applications have been running on Huawei Ascend Atlas 200I DK A2, and formed a camera, reference image, ROI, result recording and rollback closed loop. The next stage will be based on a fully China-localized solution to access Chinese imaging, control and formal commercial software environments.

Why can’t developer kit verification be directly equivalent to mass production?

The goal of the developer kit is to lower the threshold for learning, development and verification, making it easier for engineers to confirm device capabilities and application routes. Mass production systems require delivery approval, supply, environmental adaptability, maintenance, safety and quality responsibilities.

Huawei Ascend's official Atlas 200I DK A2 usage process page states that the images, tools, samples and other software provided by the developer kit can only be used by individual developers and cannot be used for commercial purposes. Therefore, formal projects cannot directly copy the current developer image into a commercial deliverable, but should confirm the commercially available software package, runtime, and hardware form.

This limitation does not negate the value of algorithm and software verification on developer kits. It requires the project to add a clear commercial migration path before entering formal delivery.

Step 1: Confirm the official commercial Atlas form and operating environment

fully China-localized mass production solutions should select Atlas acceleration modules, industrial complete machines or edge devices with clear commercial conditions, and confirm:

• Operating system and version;
• Driver, firmware, CANN or NNRT version;
• Compatibility between development environment and operating environment;
• Model and application upgrade methods;
• Security updates and technical support channels;
• Storage, cooling, power and environmental requirements.

Atlas 200I A2 commercial The project needs to confirm the actual combination based on the final hardware and algorithm route.

Step 2: Connect Chinese cameras, lenses and light sources

The fully China-localized system will use Chinese industrial cameras, Chinese lenses and Chinese light sources. The website does not need to disclose the complete BOM or supplier list, but the project implementation must complete imaging verification using actual configurations.

What needs to be verified is not "the camera can produce pictures", but:

• Whether the field of view covers all detection positions;
• Whether the minimum target and defect have enough pixels;
• Can exposure, gain and white balance be fixed;
• Whether the light source can suppress reflections and shadows;
• Whether the workpiece position and angle changes are within the algorithm tolerance;
• Whether the brightness and clarity are stable during continuous operation;
• How to deal with camera disconnection, frame acquisition timeout and abnormal images.

After the imaging conditions change, the standard images, parameter adjustment samples and independent acceptance samples need to be re-collected, and the temporary thresholds in the developer kit stage cannot be used for direct release.

Step 3: Connect to Chinese I/O and workstation control

Mass production systems must map visual results to specific workpieces. Chinese I/O, PLC or controller is responsible for triggering and result interaction, and at least needs to be verified:

• Workpiece in place signal;
• The timing from triggering to frame fetching;
• The detection is completed, PASS, FAIL and UNKNOWN are output;
• Timeout, repeated triggering and missed triggering;
• Interception of non-conforming products or manual review;
• Recovery after service restart, network disconnection and camera disconnection;
• Barcode, serial number or MES record association.

If the detection results cannot correspond one-to-one with the physical workpiece, even if the single picture algorithm performs normally, a closed loop of production quality cannot be formed.

Step 4: Select algorithm route based on workstation tasks

The algorithm route should be determined by task complexity, sample number, accuracy target and production cycle time, and cannot be selected based solely on the platform's nominal computing power.

When the workstation is fixed, the position is stable, and the differences are clear, reference image, image registration and ROI local comparison can be used. When the category is complex, the pose changes greatly, or semantic classification is required, classification, object detection, or segmentation models can be further trained.

Algorithm extensions need to be completed:

1. Training model and input and output protocol confirmation;
2. Model structure and input and output inspection;
3. Target platform model conversion and deployment;
4. Application interface access;
5. Pre-processing and post-processing implementation;
6. Verify the consistency of the same set before and after migration;
7. Performance, memory, temperature and error recovery testing.

Only after completing these steps and passing the target device test can the corresponding model performance be written into the project acceptance conclusion.

Step 5: Establish real sample acceptance

Mass production acceptance requires converting business quality standards into statistical indicators. It is recommended to distinguish at least:

• OK sample;
• Missing samples;
• Wrong sample packaging;
• Offset, blocked or reversed samples;
• Blurred, overexposed, underexposed, etc. samples cannot be judged;
• Variations between batches, shifts, temperatures and allowed locations.

reference image, parameter adjustment drawings and acceptance drawings should be isolated. The result report at least provides the key defect missed detection rate, OK false rejection rate, UNKNOWN rate, single piece cycle time, sample number and test conditions. In the absence of this information, general conclusions such as "high accuracy," "stable and reliable," or "production grade" should not be used.

Step 6: Improve permissions, security and operation and maintenance

The R&D page can usually be accessed directly on the LAN. The mass production version also requires:

• Operator, engineer and administrator privileges;
• HTTPS or controlled network access;
• inspection recipe approval, locking and modification records;
• Results, image and log retention policy;
• Device backup, recovery and offline upgrade;
• Release package hashing and version rollback;
• Fault alarm and maintenance entrance.

"Local data processing" only means that the data does not need to be sent to the public cloud, which does not mean that the system is automatically secure. Account, network, log, backup and interface permissions still need to be designed separately.

Step Seven: Complete the complete system and on-site verification

After a fully China-localized system integrates cameras, optics, computing, photography and commercial environments, it should be verified at the target site:

• Real workpiece cycle;
• Production line rhythm;
• Continuous operation;
• lighting and position changes;
• Power outage, restart and abnormal recovery;
• Recipe switching;
• Result tracing;
• Manual review process.

After completing the above work, the official website status can be updated from "Purely localized solution under development" to "Purely domesticated system joint debugging completed". Whether to use "mass production" or "production release" should also be based on the project acceptance results.

FAQ

Does a pure localization plan have to disclose the complete BOM?

No need. The official website can state that the computing, imaging, control and software environments use Chinese configurations and does not disclose the list of suppliers. Implementation and maintenance still require documenting the actual configuration and version.

Does using openEuler mean that the entire system is localized?

Not equal to. The operating system is just one layer, and also deals with the computing platform, cameras, lenses, light sources, cameras, application software and field interfaces.

Can the test results on the developer kit be directly used on commercial hardware?

It cannot be directly assumed. Installation, functionality, accuracy, tempo, and stability verification should be re-completed after changes in hardware, driver, runtime, camera, and imaging conditions.

Is deep learning necessary for mass production systems?

Not required. Fixed workstations can give priority to interpretable traditional vision; for complex scenes, classification, object detection or segmentation models can be selected based on data and indicators.

Sources and explanations

• Atlas 200I DK A2 software usage statement: https://www.hiascend.com/document/detail/zh/Atlas200IDKA2DeveloperKit/23.0.RC2/qs/qs_0018.html
• Atlas 200I A2 commercial CANN installation solution: https://www.hiascend.com/document/detail/zh/canncommercial/80RC1/softwareinst/instg/instg_0006.html

Related Atlas vision inspection resources

• Fully China-localized assembly vision inspection: Huawei Ascend Atlas project case study
• Winge Technology has completed on-device adaptation of assembly vision inspection on Atlas 200I DK A2
• What engineering work is included in Atlas 200I DK A2 vision inspection adaptation?
• How does a Chinese edge platform support assembly completeness and missing-part inspection?
• How to deploy and deliver Huawei Ascend Atlas assembly vision inspection locally
• Huawei Ascend Atlas Vision Inspection