Based on Baire BL450 ARMxy series industrial edge controller, this all-in-one platform integrates image acquisition, AI inference and isolated NPN I/O control, eliminating the need for separate vision box and PLC.
Case Details
For many single-station quality inspection tasks on small and medium-sized production lines, the traditional setup combines avision controller plus a separate PLC. This dual-hardware scheme increases BOM cost, and signal synchronization and wiring debugging take up lots of engineering time.
Drawing on the mature ARM machine vision ecosystem of RK3588J, this article shares a fully deployable closed-loop AI vision inspection solution for single stations. We cover host and I/O board selection, industrial camera wiring, part-in-place trigger, AI inference, OK/NG hardware rejection output, together with local HDMI preview. All wiring rules, trigger logic and software options are explained for machine vision engineers to reuse directly in projects.
💡 This article also clarifies a common question: whether ordinary NPN DO channels can directly output hardware trigger signals for cameras, and how to select for different working conditions.
BL450 Single-Station Machine Vision Inspection Solution
Based onBeilaiARMxy series RK3588J industrial edge controller
Application Scenarios: Appearance defect inspection, presence/absence detection, label verification, OCR character recognition for plastic parts, packaging and assembly stations. Suitable for lightweight single-station quality inspection with cycle time ≥200ms. The whole system realizes a full closed loop: part arrival → image capture → AI inference → OK/NG hardware output. One controller handles image acquisition, algorithm calculation and I/O logic without a separate PLC, and supports local HDMI display preview.
1. Hardware Selection
1.1 Host & I/O Expansion Board
Host:BL450(Industrial RK3588J, 6TOPS NPU, fanless DIN-rail mounting, built-in HDMI2.1 output) Expansion I/O board:Y01 (4-ch isolated NPN input + 4-ch isolated NPN output)
DI channels on Y01 collect workpiece in-place signals from photoelectric sensors. DO channels output OK, NG and fault alarm switch signals. All channels are galvanically isolated to protect against ESD and surge noise in workshops.
Selection guide: Choose Y11 for input-only; Y22 for output-only; Y01 is preferred for closed-loop single-station inspection. Select Y95 hardware PWM expansion board if microsecond-level hardware PWM pulse is required for camera trigger and strobe light synchronization.
1.2 Industrial Camera (2 options)
Option A: USB3 Vision Camera (Recommended)
Spec: 2MP~5MP, global shutter Wiring: USB3 cable directly connected to BL450 USB3 port; camera powered by external 24V station power supply.
Option B: GigE Industrial Camera
Spec: 2MP~5MP, global shutter Wiring: Ethernet cable to the dedicated Gigabit port of BL450; camera powered by external 24V supply. Isolate the camera subnet from office network to avoid broadcast interference.
1.3 Station Peripheral Wiring (Y01 NPN)
|
|
|
|
|
|
|
Workpiece In-place Signal
|
|
Triggers inspection workflow when workpiece arrives
|
|
|
|
Indicator / Release Mechanism
|
Isolated NPN signal for qualified products
|
|
|
|
Reject Solenoid, Alarm Lamp
|
Triggers reject action for defective products
|
|
|
|
|
Activated on camera disconnection, inference error or system failure
|
Note: Y01 DO is open-collector NPN output. External load requires 24V power supply. DO-COM must share the same ground reference with peripheral devices.
1.4 Local HDMI Display
BL450 comes with HDMI2.1 port for direct connection to monitor or industrial touchscreen.
Debug phase: Real-time preview of raw camera image, ROI selection, AI bounding boxes and OK/NG status.
Production phase: Show inspection statistics, defect snapshots and device status locally. No remote login required for on-site troubleshooting.
Tip: HDMI is mainly for debugging and on-site monitoring, not recommended as the only HMI for long-term production.
1.5 Three Camera Trigger Modes
Software Trigger (Recommended, with Y01)
Y01 DI0 captures the photoelectric in-place signal. The application receives the DI event and sends capture command to the camera via API. Advantages: Simple wiring, no jitter from IO level flip. Suitable for most stations with cycle ≥200ms.
Regular DO Hardware Pulse (Y01, Low-speed only)
Route one Y01 DO channel to camera Trigger-IN, configure camera for falling-edge hardware trigger.
Limitation: DO level toggling is controlled by software. Even with PREEMPT-RT kernel, millisecond jitter exists. Only for slow inspection with cycle ≥500ms, no flash sync requirement. Not suitable for high-speed production lines.
Hardware Timer PWM Trigger (Y95 Expansion Board, High-precision Scenarios)
Y95 generates NPN pulses via hardware PWM to camera trigger input, and synchronous strobe signal for light source. Hardware timer ensures microsecond-level jitter, ideal for strict camera and flash synchronization.
2. Software Stack Options
BL450 runs Ubuntu-RT real-time OS. Leveraging the mature RK3588J ARM vision ecosystem, two development routes are available.
Route 1: OpenCV + RKNN-Toolkit2 (Self-developed algorithm, cost-optimized)
-
OpenCV for image preprocessing: cropping, ROI extraction, filtering, distortion correction, contour filtering.
-
RKNN-Toolkit2 quantifies YOLO, segmentation and OCR models to INT8, and runs AI defect inference on the built-in NPU of RK3588J.
-
Application reads DI signals from Y01, triggers capture and preprocessing, runs AI inference, and controls Y01 DO for OK/NG output based on results.
-
Render image and detection overlay to HDMI display locally.
-
Supports Modbus-TCP / MQTT. Inspection results and defect images can be uploaded to MES for production traceability.
Route 2: Hikvision VisionMaster-ARM Runtime (Fast deployment, minimal algorithm development)
Build inspection workflow on Windows with VisionMaster GUI, export project and deploy to BL450 Runtime. Read DI signals to launch inspection tasks and control I/O output via underlying API. Detection image and judgment result can also be output to HDMI.
Supplementary Notes: Halcon Linux-AArch64 Runtime works, but ARM licenses are premium. Traditional template and caliper operators perform worse than on X86 platforms. Cognex VisionPro has no ARM version and is unsupported in this solution.
Optional Functions
Codesys SoftPLC can be pre-installed to combine vision results with PLC logic. Docker container deployment is supported.
3. Complete Workflow
-
Workpiece arrives at the inspection station. NPN photoelectric sensor sends in-place signal to Y01 DI0.
-
The controller receives the signal and triggers the industrial camera to capture an image.
-
OpenCV preprocesses the image, then the RK3588J NPU runs AI defect inference.
-
Judgment result drives hardware output, meanwhile the image and detection result are shown on HDMI:
OK: Y01 DO0 outputs NPN signal for release.
NG: Y01 DO1 outputs NPN signal to activate the reject solenoid, and defect image is saved.
Inspection data is uploaded to MES on demand. The system waits for the next workpiece and cycles.
4. Problems Solved by This Solution
Simplify hardware architecture and cut BOM cost
Traditional schemes require separate vision box and PLC. BL450 integrates image capture, AI inspection and isolated I/O control in one unit. Less hardware, simpler wiring and no cross-device synchronization debugging. Great for labellers, sorting machines and small custom assembly machines.
Local HDMI visualization for convenient debugging and monitoring
HDMI output allows live preview of images and detection marks during commissioning. Operators can check status directly on-site without remote access.
Lower algorithm licensing cost for small & medium automation integrators
OpenCV + RKNN open-source stack avoids recurring runtime license fees charged by commercial vision libraries per station.
Isolated industrial I/O for reliable workshop operation
Y01 DI/DO are fully galvanically isolated. OK/NG is hard switch output. Reject and alarm actions remain reliable even if network communication fails.
Compact fanless DIN-rail design for tight control cabinets
Small form factor, fanless, wide temperature range for factory environments.
Good scalability for future production upgrades
Add PCIe computing stick for higher AI performance; stack additional Y-series I/O boards for more channels; connect to MES for data storage and traceability.
5. Scope & Limitations
-
Software trigger is recommended for cycle ≥200ms. Regular DO trigger is only for cycle ≥500ms slow sampling. Y95 PWM board is required for camera-light synchronization.
-
Optimized for qualitative inspection: presence check, defect classification. Not suitable for high-speed precision dimensional measurement with massive sub-pixel caliper and template matching.
-
VisionPro is unavailable on ARM. Evaluate operator performance and licensing cost before adopting Halcon ARM runtime.
-
HDMI is for debugging and monitoring only, not intended as primary HMI for long-running production.
Closing
If you are building lightweight single-station AI vision projects, this hardware-software-I/O closed-loop solution is worth reference. Feel free to leave comments to share your experience and pitfalls of ARM machine vision deployment.
