(Note: This project serves as a PLC-Python AI bridge designed to work in integration with the AI-Powered Fabric Defect Marking System and Industrial CANbus Data Analyzer projects.)
This project is an Artificial Intelligence (AI) based control system that detects defective products/surface anomalies on the production line and sends real-time signals to a SIEMENS PLC (S7-1200 / S7-1500) using the Snap7 library.
Update (v2.0): The previous color-based (OpenCV HSV Thresholding) Proof-of-Concept algorithm has been completely replaced with a YOLOv8 object detection model trained on the MVTec AD dataset, meeting real-world industrial standards.
- Webcam / Video Stream: Provides real-time video feed from the production line.
- YOLOv8 AI Analysis: Detects flaws and anomalies on industrial surfaces.
- PLC Communication: Python automatically writes Data Blocks (DB) to the PLC via Snap7. (Sends
2if a defect is found,1if the surface is OK.) - Real-Time Decision: Automatically triggers the connected PLC automation system to stop the line when a defect is detected.
Install the core libraries required to run the project:
pip install python-snap7 opencv-python ultralyticsTo operate properly, the YOLOv8 model must be trained using the MVTec AD dataset.
- Download Dataset: Download a specific category (e.g.,
leather,wood,metal_nut) from the MVTec AD Dataset. - Extract Archive: Extract the Zip/Tar file into the
mvtec_anomaly_detection/folder in the project's root directory. (e.g., directory structure should bemvtec_anomaly_detection/leather/trainandtest). - Convert to YOLO Format: Run the following command in your terminal to convert the dataset masks into YOLO Bounding Box format:
python src/prepare_mvtec.py --category leather(This process will generate a datasets/mvtec_leather_yolo folder and a mvtec_leather_data.yaml config file.)
To train the model dynamically on your local machine:
python src/train_yolov8.py --data ./mvtec_leather_data.yaml --epochs 50(Training duration varies depending on your hardware. Once complete, results are saved in the runs/detect/mvtec_surface_inspection folder, and the best weights will be stored as weights/best.pt.)
Open the src/ai_inspection.py file and update the PLC_IP, RACK, SLOT, and DB_NUMBER variables to match your physical or simulated PLC environment.
After the training is completed (or using the default yolov8n.pt fallback), start the main inspection system:
python src/ai_inspection.pyThe mAP scores and performance graphs obtained as a result of the training are listed below. The model was trained for 50 epochs using the YOLOv8 Nano architecture to optimize the training duration.
- Model: YOLOv8 Nano (
yolov8n.pt) - Dataset: MVTec AD (
leathercategory) - mAP50: 79.7%
- mAP50-95: 47.4%
Detailed training loss graphs and validation batch samples can be found inside the runs/detect/ directory of your project space.
This proposal introduces an "Image Processing-Based Surface Inspection System."
The system detects defects on manufactured rollers using AI, communicates those results directly to the Siemens PLC infrastructure, and manages production in real time — without manual operator intervention.
Current quality control relies entirely on operator attention, which introduces:
| Risk | Impact |
|---|---|
| Eye fatigue → missed defects | Defective products reaching customers |
| No historical data logging | Root-cause analysis is impossible |
| Inconsistent inspection speed | Quality variance between operators / shifts |
An automated inspection station using high-resolution industrial cameras and AI algorithms.
┌──────────────────────────────────────────────────────────────┐
│ [Industrial Camera] │
│ │ │
│ ▼ │
│ [PC / Python Server] ──── S7 Protocol (Snap7) ────► [PLC]│
│ YOLOv8 / OpenCV S7-1200 │
│ Defect Classification TIA V17 │
│ │ │ │
│ ▼ ▼ │
│ [HMI Display] [Tower Lamp]│
│ Defect type + location OK / FAULT │
└──────────────────────────────────────────────────────────────┘
Layer 1 — Vision (PC/Server):
- Python analyzes the moving product image using open-source libraries
- Classifies defects: scratch, hole, stain
Layer 2 — Communication (S7 Protocol via Snap7):
- The PC communicates with the PLC using Siemens's native S7 Communication protocol
- The Python software writes the analysis result (OK / NOK) directly into PLC memory (Data Block) — faster and more reliable than Modbus
Layer 3 — Action (PLC — TIA Portal):
- When
STATUS = FAULT(2), the PLC brings the machine to a safe stop - Defect type is displayed on the HMI panel
- Audible/visual tower lamp alerts the operator
| Component | Specification |
|---|---|
| PLC | Siemens S7-1200 / S7-1500 Series |
| PLC Software | TIA Portal V16/V17 (SCL + Ladder) |
| Vision Language | Python 3.9 (OpenCV, python-snap7, NumPy) |
| AI Model | YOLOv8 (v2.0 upgrade from color-based PoC) |
| Protocol | S7 Communication — Ethernet/TCP Direct Put/Get |
| Benefit | Details |
|---|---|
| Near-Zero Defect Escape | ~100% quality assurance on outbound products |
| Efficiency | Early defect detection prevents raw material and time waste |
| Data-Driven | Every defect is logged with timestamp and image for traceability |
| Week | Milestone | Status |
|---|---|---|
| 1 | System design and simulation | ✅ Done |
| 2 | TIA Portal SCL programming | ✅ Done |
| 3 | Physical PLC + Snap7 integration | ✅ Done |
| 4 | Final testing and presentation | ✅ Done |
| 5 | Replace Color PoC with trained YOLOv8 model | ✅ Done |
| Metric | Value |
|---|---|
| Total response time (camera → PLC signal) | < 100 ms |
| Deep learning model accuracy (YOLOv8) | ~80% mAP50 (Leather defects) |
| PLC reaction time (trigger → machine stop) | < 10 ms (within one PLC scan cycle) |
| Communication stability | Direct DB memory access via Snap7 — no middleware delays |