Wenge Intelligent Manufacturing Wenge Intelligent ManufacturingIndustrial smoke detection and edge AI deployment
简体中文English日本語한국어
Submit Similar Need
Home/Cases
AI Vision Delivery Case

RK3588 Industrial Smoke and Black-Smoke Detection System Delivery

This case covers industrial smoke detection and black-smoke alarm requirements, including edge smoke detection, black-smoke color classification, RTSP camera integration, MCP2221A GP0 output linkage, false-positive guards, RK3588 industrial-computer deployment and site verification. Public content removes customer network addresses, device accounts, remote access and internal paths. The black-smoke class is a project algorithm category, not a Ringelmann grade or regulatory emission conclusion.

RK3588 edge deploymentTwo-stage smoke and black-smoke judgementRTSP stream integrationMCP2221A GP0 linkage

Case Facts

Project type
Edge smoke and black-smoke detection system deployment on an RK3588 industrial computer.
Stack
Python web service, Nginx/Gunicorn, OpenCV/ONNX, MobileNetV3-Small, RTSP camera and MCP2221A GPIO.
Recognition path
A three-frame temporal model first detects smoke, then daylight frames enter black/non-black smoke color classification. Night color classification is disabled.
Output linkage
When smoke, black color confidence and hold-time conditions are met, MCP2221A GP0 outputs the configured alarm level and returns to automatic control after the test.
Sample scope
The sample package contained 59 main training images, 109 YOLO smoke boxes, color-classification crops, segmentation masks, source licenses and SHA-256 manifests. It did not include customer videos or device configuration.
Verification boundary
After v2.0.3 deployment, service and model status were checked and 70 automated tests passed. Results depend on camera angle, exposure, weather, samples and site acceptance criteria.

Project Background

The customer needed a single camera to observe smoke changes at an industrial site and output an electrical signal when black-smoke alarm conditions were confirmed. The system had to run locally on an RK3588 industrial computer to reduce dependence on external networks and cloud inference.

Industrial smoke monitoringLocal inference and console previewAlarm-level output linkage

Algorithm and System Implementation

The system uses a two-stage path. A three-frame MobileNetV3-Small ONNX model first detects smoke. During daylight, a second-stage color classifier distinguishes black smoke from non-black smoke. The server reads RTSP, HTTP or local video sources through OpenCV and FFmpeg and converts them to browser-previewable MJPEG.

Three-frame temporal smoke classificationDaylight black-smoke color classificationBrowser console preview

False-Positive Guards and Regression

The current model configuration is smoke_acceptance_model_v2.6. It includes fixed-camera clear-background protection, structure-constrained static-scene gating, pure-black/pure-color/near-binary graphic guards and an overall_dark guard for extremely dark abnormal frames. In v2.0.3, the new rule was checked against 1,760 readable images without hitting real samples, and 70 automated tests passed.

Fixed-camera background guardStatic-scene gateExtremely dark frame guard

Site Deployment and Output Linkage

In v2.0.0, image and video tests were verified with MCP2221A output linkage. A single black-smoke test image produced 98.53 percent smoke confidence and 99.99 percent color confidence. A video test sampled 74 frames: 74 were classified as smoke and 73 entered black-smoke candidates. GP0 output high during the test and returned low after the hold period. These results apply only to this test material and site configuration.

Image and video test linkageGP0 level verificationReturn to automatic control after test

Video Stream and Camera Integration

During site troubleshooting, the original 4K H.265 main stream reproduced continuous read failures after about 36 seconds. After switching to a 1080p H.264 substream at about 10 FPS, the MJPEG preview output 235 frames in 40 seconds and remained connected. The finding applies to this camera encoding and network condition.

RTSP camera integrationH.264 substream adjustmentMJPEG preview path check

Applicable Scenarios and Boundaries

This pattern fits fixed-camera industrial smoke observation, edge AI warning, LAN console preview and GPIO alarm output projects. The black-smoke result is a project algorithm output and does not replace official environmental monitoring, emission-grade measurement or regulatory evidence. Formal acceptance should cover no-smoke, smoke onset, continuous black smoke, stream loss, reboot, lens occlusion, rain/fog and backlight conditions.

Fixed-camera industrial visionEdge AI alarm outputSite acceptance criteria first

Need industrial smoke detection or edge AI alarm assessment?

Submit site videos, device list, output interfaces and acceptance targets so the algorithm class, deployment method, rollback plan and validation path can be confirmed first.

Request Assessment