Industrial Anomaly Detection

Overall block diagram

User browser
    ↓ HTTP :5000
┌────────────────────┐
│  Flask Web          │
│  app.py             │
└────────┬───────────┘
         ↓
┌────────────────────┐
│  PatchCore engine   │
│  patchcore_train.py │  ← training (CPU builds the memory bank)
│  patchcore_test.py  │  ← inference
└────────┬───────────┘
         ↓
┌────────────────────┐
│  RKNN3 Lite         │  ← rknn3-toolkit-lite 1.0.0 installed
│  DINOv3-S backbone  │
└────────────────────┘

The board currently has Flask 3.1.3 (Debian 12 apt's python3-flask); PatchCore's ML dependencies (numpy / scipy / torch, etc.) are not installed, so run.sh start reports ModuleNotFoundError.

1. What Industrial Problems Can the Anomaly Detection Application Solve?

The RK182X anomaly detection application is based on the PatchCore algorithm + DINOv3-Small backbone, designed for industrial surface defect detection.

Actual path: /userdata/RK1820_RK1828_AI_SDK/examples/anomaly-detection/

Typical scenarios:

• Bottles: bubbles, scratches, dents
• Cables: insulation damage, deformed connectors
• Fabrics: holes, stains, weaving defects
• Metal surfaces: scratches, oxidation, corrosion

Application workflow: first train with good samples (building a feature memory bank), then run anomaly detection on test samples and output a heatmap.

2. Actual Deployment Status

Application directory:

ls /userdata/RK1820_RK1828_AI_SDK/examples/anomaly-detection/

Output:

app.py
install.sh
model
patchcore_test.py
patchcore_train.py
pictures
README_CN.md
README_EN.md
run.sh
static
system
templates
THIRD_PARTY_LICENSES.txt

Total size:

du -sh /userdata/RK1820_RK1828_AI_SDK/examples/anomaly-detection/

Output:

45M	/userdata/RK1820_RK1828_AI_SDK/examples/anomaly-detection/

Flask installation status:

python3 -c "import flask; print('Flask version:', flask.__version__)"

Output:

Flask version: 3.1.3

Flask 3.1.3 comes from Debian 12 apt's python3-flask.

Basic functionality:

python3 -c "import flask; from flask import Flask, render_template; print('Flask基础功能可用')"

Output:

Flask基础功能可用

Model files:

ls -lh /userdata/RK1820_RK1828_AI_SDK/examples/anomaly-detection/model/

Output:

总计 43M
-rw-r--r-- 1 linaro linaro 613K  2026年 1月28日 dinov3s.rknn
-rw-r--r-- 1 linaro linaro  42M  2026年 1月28日 dinov3s.weight

Python scripts:

ls -la /userdata/RK1820_RK1828_AI_SDK/examples/anomaly-detection/*.py

Output:

-rw-r--r-- 1 linaro linaro 26823  2026年 1月28日 app.py
-rw-r--r-- 1 linaro linaro  5726  2026年 1月28日 patchcore_test.py
-rw-r--r-- 1 linaro linaro 12168  2026年 1月28日 patchcore_train.py

Patchcore dependencies:

cd /userdata/RK1820_RK1828_AI_SDK/examples/anomaly-detection/system/install && unzip -p patchcore-0.1.0-py3-none-any.whl patchcore-0.1.0.dist-info/METADATA | grep -A 10 "Requires-Dist"

Output:

Requires-Dist: click (>=8.0.3)
Requires-Dist: faiss-cpu
Requires-Dist: matplotlib (>=3.5.0)
Requires-Dist: pillow (>=8.4.0)
Requires-Dist: pretrainedmodels (>=0.7.4)
Requires-Dist: scikit-image (>=0.18.3)
Requires-Dist: scikit-learn (>=1.0.1)
Requires-Dist: scipy (>=1.7.1)
Requires-Dist: timm
Requires-Dist: torch (>=1.10.0)
Requires-Dist: torchvision (>=0.11.1)
Requires-Dist: tqdm (>=4.62.3)

Installed related packages:

python3 -m pip list | grep -E "(click|scipy|tqdm)"

Output:

click                        8.4.2
tqdm                         4.70.0

scipy is not installed — install.sh only installs python3-flask + patchcore-0.1.0-py3-none-any.whl; it does not install ML libraries such as scipy / numpy / torch / timm.

3. How to Deploy?

3.1 Push Files to the Board

# On the PC
adb push examples/anomaly-detection /userdata/

3.2 Prepare the Dataset

Download the MVTec AD dataset:

https://www.mvtec.com/company/research/datasets/mvtec-ad

Directory structure:

/userdata/datasets/
├── bottle/
│   ├── test/
│   │   ├── good/         ← good test samples
│   │   └── defect/       ← defective test samples (one subdirectory per defect type)
│   ├── train/good/       ← training uses only good samples
│   └── ground_truth/     ← defect annotations
├── cable/
│   ├── test/
│   └── train/
└── ...

3.3 Install Dependencies

Startup script:

cat /userdata/RK1820_RK1828_AI_SDK/examples/anomaly-detection/run.sh | head -15

Output:

#!/bin/bash -i

start() {
    start-stop-daemon --start \
        --pidfile /run/anomaly-detection.pid --make-pidfile \
        --chdir /userdata/anomaly-detection \
        --exec /usr/bin/python3 -- /userdata/anomaly-detection/app.py --webbrowser
}

The script points to /userdata/anomaly-detection, but the actual path is /userdata/RK1820_RK1828_AI_SDK/examples/anomaly-detection.

Installation:

cd /userdata/RK1820_RK1828_AI_SDK/examples/anomaly-detection
./install.sh

install.sh installs:

• python3-pip
• python3-flask
• patchcore-0.1.0-py3-none-any.whl

Important limitation: full operation requires many machine-learning dependency libraries (torch, torchvision, faiss-cpu, timm, pretrainedmodels, scikit-image, scikit-learn, numpy, etc.); these are hard to install on edge devices and consume substantial resources.

3.4 Start the Application

# Method 1: start from a terminal
/userdata/anomaly-detection/run.sh start

# Method 2: desktop icon (created by install.sh)
# Click "工业检测" (Industrial Inspection) on the desktop

# LAN access
http://<board IP>:5000

4. Workflow

4.1 Labeling (Training)

1. Select a dataset in the web interface
2. Click the "标注" (Label) button; the system extracts good-sample features with the DINOv3 backbone
3. After labeling completes, the anomaly threshold is computed

4.2 Testing

1. Select a labeled dataset on the "数据集测试" (Dataset Test) page
2. Click "开始测试" (Start Test)
3. Results are shown as heatmaps, green = normal, red = anomalous

Test page buttons:

Button	Function
Dropdown	Select the dataset
自动运行 (Auto run)	Whether to run continuously (pauses on anomaly by default)
运行线程数 (Threads)	Set the number of parallel threads
开始测试 (Start Test)	Start the test
停止运行 (Stop)	Stop the current test
连续运行 (Run all)	Detect every image without pausing
运行至异常 (Run to anomaly)	Continue after an anomaly pause
下一张 (Next)	Advance one image per thread
热力图 (Heatmap)	Click the eye icon at the top-right of an image

5. Actual Directory Structure

examples/anomaly-detection/
├── app.py                     # Flask web app (26KB)
├── patchcore_train.py         # PatchCore training (12KB)
├── patchcore_test.py          # PatchCore testing (5.7KB)
├── install.sh                 # Installation script
├── run.sh                     # Start/stop script
├── model/
│   ├── dinov3s.rknn           # 613KB (DINOv3-S structure)
│   └── dinov3s.weight         # 42MB (DINOv3-S weights)
├── README_CN.md / README_EN.md
├── pictures/                  # Operation screenshots
│   ├── step1_cn.jpg          # Chinese step screenshots
│   ├── step1_en.jpg          # English step screenshots
│   └── ...
├── static/                    # Flask static assets
│   ├── css/
│   ├── js/
│   ├── fontawesome-free-7.0.0-web/
│   ├── logo.png
│   └── datasets -> ../../datasets  # symlink to the datasets
├── system/
│   ├── install/
│   │   ├── patchcore-0.1.0-py3-none-any.whl
│   │   └── pip.conf
│   ├── desktop/               # Desktop icon entry
│   └── lib/
└── templates/                 # Flask HTML templates
    └── index.html            # Main page (6.9KB)

Static assets:

ls /userdata/RK1820_RK1828_AI_SDK/examples/anomaly-detection/static/

Output:

css
datasets
fontawesome-free-7.0.0-web
js
logo.png

Templates:

ls -la /userdata/RK1820_RK1828_AI_SDK/examples/anomaly-detection/templates/

Output:

-rw-r--r-- 1 linaro linaro 6983  6月 8日 14:28 index.html

6. Prerequisites

The RK182X module is connected (PCIe):

systemctl is-active rknn3.service

Output:

active

ls -l /dev/pcie-rkep-*

Output:

crw------- 1 root root 10, 124  8月20日 16:56 /dev/pcie-rkep-0004:41:00.0

The MVTec AD dataset must be downloaded to /userdata/datasets/ (after actual deployment this directory does not exist yet; download and extract MVTec manually):

ls /userdata/datasets/

Important limitation: full operation requires many machine-learning libraries (torch, torchvision, faiss-cpu, numpy, scipy, etc.); these are hard to install on edge devices and consume substantial resources (verified: none are installed).

7. FAQ

Symptom	Cause	Fix
run.sh start reports ModuleNotFoundError: No module named 'numpy'	scipy/numpy/torch not installed	install.sh does not install ML libs; pip install manually
Flask fails with Address already in use	Port 5000 occupied	Find the occupier with ss -tlnp | grep 5000
dinov3s.rknn fails to load	rknn3-toolkit-lite not installed	apt install rknn3-toolkit-lite
run.sh path not found	Points to /userdata/anomaly-detection, actually in the SDK path	Fix the path or create a symlink
datasets symlink broken	/userdata/datasets not created	Create it manually + download MVTec AD
MVTec dataset missing some of the 15 classes	Incomplete download	Re-download the full tarball

8. Next Steps

• ClawChips Architecture and Principles — ModelHub scheduling
• SKILL User Manual — the rk-inspect inspection Skill
• AI Agent Applications — OpenClaw integration