← Back to case studies Industrial AI case

Industrial black-smoke detection and alarm system: from video analytics to RK3588 edge deployment

An anonymized Sinovanta case covering two-stage smoke detection, black- and white-smoke classification, RK3588 edge deployment, GPIO alarm control, and evidence-based validation.

Engagement
Industrial vision and edge AI
Scope
Smoke detection, color classification, and alarm output
Platform
IP camera, RK3588, and MCP2221A GPIO
Verified release
Software v2.0.3 with model configuration v2.6
01

Project context

An industrial site required a camera-based system to distinguish smoke from non-smoke scenes and then classify detected smoke as black, white, or uncertain. The solution also needed to run on an RK3588 industrial computer, accept an RTSP camera stream, provide browser-based monitoring and media tests, and drive a physical alarm output.

02

Delivery objective

Deliver a traceable edge system covering data preparation, two-stage inference, false-positive controls, web operation, hardware output, deployment, rollback, and acceptance records without treating a single video result as general model accuracy.

Two-stage edge inference and alarm control

The alarm path separates smoke presence from smoke color and adds temporal confirmation before any hardware output. This keeps the decision chain inspectable and allows false-positive protections to be tested independently.

  1. 01

    Video ingestion

    Receive the IP-camera RTSP stream or an uploaded image or video and normalize frames for the configured analysis mode.

  2. 02

    Smoke detection

    The first stage determines whether smoke is present. Non-smoke artifact guards reject pure-black, near-binary, static, and abnormally dark inputs before alarm logic.

  3. 03

    Smoke-color classification

    Only confirmed smoke enters the second stage, which classifies black smoke, white smoke, or uncertain color using scene-aware sampling.

  4. 04

    Temporal confirmation

    Live and uploaded video results must meet the configured confidence and persistence conditions before an alarm request is created.

  5. 05

    Physical output

    The RK3588 service controls MCP2221A GP0, holds the configured alarm level, and then returns to automatic monitoring.

  6. 06

    Field hardening

    Regression tests, device-side checks, service health, stream state, file hashes, backups, and rollback records are kept with each release.

03

Controlled deliverables

Algorithm and model configuration
Two-stage smoke and color logic, scene guards, thresholds, and versioned model configuration.
RK3588 web application
Live monitoring, image and video testing, settings, status, and controlled alarm-link options.
Hardware integration
MCP2221A connection, GP0 output control, hold timing, recovery to automatic mode, and output-error reporting.
Deployment package
Service files, installer, checksums, startup configuration, backup and rollback records.
Validation records
Dataset manifest, automated tests, device-side evidence, regression results, and documented limitations.
04

Verified results

The figures below are tied to named test inputs, software versions, and device conditions. They are not presented as general accuracy claims.

37-second test video on RK3588

74 of 74 sampled frames detected smoke; 73 of 74 were classified as black smoke; black-smoke share was 98.65%; the alarm condition was met.

This evidence run used isolated upload output, so it did not actuate GPIO.

Hardware-linked media test

A qualifying black-smoke image and video drove GP0 high, held the output for 11 seconds, and then returned to automatic mode with GP0 low.

Verified on the customer industrial computer in software v2.0.0.

Dark-input false-positive regression

Thirteen pure-black, near-black, noisy, gradient, JPEG, PNG, and text variants produced no alarm; three device-side noisy or gradient inputs produced no GP0 high.

Verified in software v2.0.3 with the v2.6 model configuration.

Dataset and automated regression

The new overall-dark guard blocked 0 of 1,760 readable existing dataset images; 70 automated tests passed.

The scan supports this guard only for the current dataset and thresholds.

05

Acceptance and operating evidence

  • Deployed application and model files matched recorded SHA-256 values.
  • The service was active and enabled, the RTSP stream was connected, and MCP2221A was available during the recorded acceptance checks.
  • A real black-smoke benchmark remained detected at 98.53% smoke confidence and 99.99% color confidence after the dark-input guard was added.
  • Six post-test live samples showed no smoke, no active alarm, GP0 low, and the hardware interface connected.
06

Scope and limitations

These results apply to the documented videos, images, dataset, camera conditions, RK3588 device, model configuration, and thresholds. They do not establish general precision, recall, certified opacity measurement, emissions-compliance monitoring, or continuous 24-hour availability. A new site requires camera-position review, representative day and night data, threshold calibration, obstruction and weather tests, output wiring verification, and a separately agreed endurance test.

Discuss an industrial vision requirement

Provide the camera source, target scene, alarm logic, edge platform, available media, operating conditions, hardware interface, and acceptance criteria.

Submit your requirements →