Published: 2026-07-27 | English reference version of the Chinese source article.
Industrial AI starts with the systems already in place
Industrial parks, factories, warehouses and public sites have spent years building cameras, switches, NVRs and management platforms. Most of those systems still centre on recording and manual review. A missed helmet, a person in a restricted area, a blocked route or smoke and flame can remain invisible until an operator sees the related view.
An edge approach puts compute at the site. The device takes streams from cameras or an NVR, decodes them, runs a model, applies a rule, keeps evidence and outputs an event. Existing cameras keep their imaging role and the NVR keeps its recording role, so a project can upgrade incrementally.
Product and interface scope
XINHUO AI Edge Box product material describes RTSP, ONVIF and GB28181 stream access, local tasks for PPE, behaviour, smoke and flame, vehicles and region rules, plus Web, HTTP and MQTT alarm output. The product line covers 2, 8, 16 and 32 channel configurations with a Web console and operating-state monitoring.
Public product specifications describe design and interface capability. Public procurement records describe whether a listed device appeared in a particular procurement. They answer different questions and should not be used as substitutes for a project's own accuracy test or acceptance record.
The difference between a model result and an operational event
A vehicle detector can identify a vehicle. A parking violation rule also needs to decide whether the vehicle is inside a defined area, whether the dwell time has passed a threshold, whether the same case was already reported and whether the evidence is clear enough for review.
Similar rules are needed for leaving a post, sleeping on duty, intrusion, fence climbing and accumulated objects. The delivery record normally includes the device, channel, time, algorithm type, captured evidence and target attributes. Operations staff use those records alongside device load, task state and delivery time to identify whether a fault belongs to the camera, network, task or platform connection.
Hardware selection is a field decision
The product range includes a 2 TOPS INT8 two-channel unit, an eight-channel XH-A30-B1 with an eight-core CPU, 6 TOPS INT8 NPU, 8 GB memory and 64 GB storage, 16-channel configurations, and a 32-channel video analytics server with a 20-core 40-thread CPU, GPU acceleration and 32 GB memory. Actual concurrent capacity still follows the selected streams, algorithms, preview load and delivery rules.
The published material identifies domestic AI chip options across the product range, including Huawei Ascend for a 16-channel 20 TOPS configuration, Rockchip RK3576 for a 6 TOPS configuration, and Allwinner A733 for a 3 TOPS configuration. Hardware details need to be confirmed against the model delivered to the project.
Public application records cited in the Chinese source
The source article cites a June 5, 2024 campus security upgrade result notice from Zhuhai City Polytechnic, a May 26, 2026 award notice for a refined-management project in Beijing's Fengtai vehicle-repair sector, and a product record from China Huaneng Group's marketplace. The notices identify particular procurement, model or product-record facts. They do not publish all deployment conditions or results for other projects.
A sustainable deployment still needs a site survey, representative video, model and rule testing, interface confirmation and an acceptance record. That is where a camera view, the event workflow and the operator's response become one operating system.