Smart Industrial Vision Edge Computing Terminal Technical Solution Document‌

‌1. Product Overview‌

1.1 Product Positioning

High-performance AI vision processing equipment for industrial automation scenarios
Integrated BM1684X (8TOPS INT8) AI acceleration chip + RK3588 (6TOPS NPU) heterogeneous computing
Supports industrial-grade multi-camera access, real-time defect detection, high-precision positioning, OCR recognition, and other scenarios

1.2 Core Advantages

‌Dual-chip Synergy‌: BM1684X focuses on deep learning inference, while RK3588 handles multi-channel video encoding/decoding and system control.
‌Industrial-grade Reliability‌: Wide temperature design (-20℃~70℃), EMI resistance, supports 24/7 continuous operation.
‌Algorithm as a Service‌: Pre-installed with 20+ industrial vision algorithm models, supports one-click deployment of TensorFlow/PyTorch/Caffe models.

‌2. Hardware Specifications‌

2.1 Processor Unit

Module | Parameter Description ---|--- AI Acceleration Chip | Sophon BM1684X, 8TOPS INT8 computing power, supports INT4/FP16/BF16 Main Control Processor | Rockchip RK3588, Quad-core A76 + Quad-core A55, 6TOPS NPU

2.2 Vision Interfaces

‌Camera Access‌: 2x GigE Vision PoE+, 4x USB3.0 Vision (compatible with industrial cameras like Basler/Hikvision)
‌Display Output‌: HDMI 2.1 (4K@60Hz), LVDS industrial display interface
‌Optical Support‌: External trigger signal input (5-24V), strobe control output

2.3 Expansion and Communication

‌Industrial Bus‌: 2x CAN 2.0B, 1x RS485 (isolated protection)
‌Network‌: Dual Gigabit RJ45 (supports EtherCAT master protocol)
‌Wireless‌: Optional WiFi6/5G module

2.4 Mechanical and Environmental

‌Protection Rating‌: IP65 all-metal enclosure, fanless cooling design
‌Mounting Method‌: DIN rail mounting + wall mounting
‌Power Input‌: 12-36V DC wide voltage input, overvoltage/reverse polarity protection

‌3. Software Architecture‌

3.1 System Layer

‌Operating System‌: Ubuntu 20.04 LTS (real-time kernel patch)
‌Container Support‌: Docker engine, supports containerized algorithm deployment

3.2 Development Kit

‌SDK‌: Python/C++ API, provides full-process interfaces for camera control, pre-processing, and post-inference processing.
‌Toolchain‌: Sophon TPU-MLIR model compiler, RKNN-Toolkit2 model conversion tool.

3.3 Algorithm Repository (Pre-installed)

Algorithm Type | Typical Functions ---|--- Defect Detection | Surface scratches, solder joint defects, foreign object detection (detection rate ≥99.5%) Positioning and Guidance | Sub-pixel template matching (accuracy ±0.02mm) Measurement | Geometric dimension measurement, angle calculation (repeatability ±1μm) Recognition | QR code/OCR recognition (supports multi-language mixed layout)

‌4. Typical Application Scenarios‌

4.1 Electronics Manufacturing Industry

SMT component positioning
PCB solder joint quality inspection
Chip Mark point recognition

4.2 Metal Processing

Precision part dimension measurement
Metal surface oxide layer detection
Stamped part burr recognition

4.3 Packaging Industry

Print quality inspection
Packaging integrity inspection
Production date OCR verification

‌5. Performance Metrics‌

Item | Parameter ---|--- Inference Frame Rate | 1080P@30fps (YOLOv5s model) System Latency | Image input to IO output ≤50ms Multi-task Capability | Simultaneously runs 4 independent vision tasks Model Switching Time | <5s (hot switching without reboot) MTBF | ≥50,000 hours