92.95% mAP50 Valve Detection + 73.06% mAP50 Anomaly Detection

ValveAI

End-to-end AI inspection pipeline for underground gas valve wells. Detect 4 valve types, identify 6 anomaly classes (corrosion, cracks, water damage, coating damage, fog), and generate health scores โ€” across 861,000+ inspection photos. Available in PyTorch, ONNX, CoreML, and TorchScript formats.

Try Live Demo Get Model View Source
92.95%
Valve Detection mAP50
861K
Images Analyzed
6
Anomaly Classes
73.06%
Anomaly Detection mAP50 (V13)
4
Model Formats

Real-Time Valve Detection

Valve Detection Demo
๐Ÿ”ฌ
4 Valve Classes
Gate valve, Globe valve, Ball valve, and Other valve types โ€” covering all common underground gas infrastructure.
โšก
Real-Time Inference
~8ms per image on CPU, ~3ms on GPU. Process thousands of inspection photos in minutes.
๐Ÿ“
GPS Integration
93% of images have GPS coordinates. Generate interactive inspection maps with 650+ geolocated sites.
๐Ÿ”ง
Production Ready
PyTorch, ONNX, CoreML, TorchScript exports. Edge deployment ready. REST API endpoint. Ready for integration.

End-to-End Inspection Pipeline

3 AI models working together: valve detection + anomaly detection + anomaly classification = health score

1
Valve Detection
YOLOv8s
92.95% mAP50
4 valve types
2
Anomaly Detection
YOLOv8s
73.06% mAP50
6 anomaly bboxes
3
Anomaly Classification
EfficientNet-B0
74.0% accuracy
severity levels
=
Health Score
0-100 score
GOOD / FAIR / POOR / CRITICAL
inspection report

10 Rounds of Iterative Training

From 30 labeled images to 92.95% accuracy โ€” without paying for annotations

28%
R1
30 imgs
66%
R4
1.6K imgs
81%
R7
3.9K imgs
84%
R8
8.5K imgs
81%
R9
18.6K imgs
92.95%
R10
9K imgs

Key Insight: R9 added more data but hurt performance due to noisy labels. R10 filtered to high-confidence labels only โ†’ +6.9% mAP50 improvement.

Anomaly Detection System

AI-powered defect identification across 861,000+ underground valve well inspection photos

๐Ÿฆ 
Corrosion & Rust
่…่š€็”Ÿ้”ˆ
77.3% prevalence
๐ŸŽจ
Coating Damage
ๆถ‚ๅฑ‚ๆŸๅ
67.7% prevalence
๐Ÿ“
Wall Cracks
ๅข™ๅฃ่ฃ‚็ผ
59.6% prevalence
๐Ÿ’ง
Water Accumulation
็งฏๆฐด
44.4% prevalence
๐ŸŒง๏ธ
Water Seepage
ๆธ—ๆฐด
46.1% prevalence
๐ŸŒซ๏ธ
Fog Condensation
้›พๆฐ”ๅ‡็ป“
11.0% prevalence
94.5%
Images with anomaly
68.3%
Mild severity
73.06%
Anomaly mAP50 (V13)
EfficientNet-B0
Classifier (74% accuracy)
YOLOv8s V13
Bbox detection (+15.5% vs V2s)
Get Anomaly Detection Model

4 Valve Classes

๐Ÿ”ง
Gate Valve
้—ธ้˜€
~45% of data
๐Ÿ”ฉ
Globe Valve
ๆˆชๆญข้˜€
~17% of data
โš™๏ธ
Ball Valve
็ƒ้˜€
~34% of data
๐Ÿ”ต
Other Valve
ๅ…ถไป–
~4% of data

Professional Dataset Available

9,038
Anonymized Images (EXIF stripped)
45,134
Bounding Box Annotations
1.8 GB
YOLO Format, Ready to Train
Browse Dataset on Hugging Face

Licensing & Services

Valve Detection Only
$499
one-time
  • YOLOv8s valve detection (92.95% mAP50)
  • ONNX export for deployment
  • Python inference script
  • Commercial usage license
  • Documentation & examples
Complete Pipeline
$1,499
one-time
  • Valve detection model (92.95% mAP50)
  • Anomaly detection model V13 (73.06% mAP50, 6 classes)
  • EfficientNet-B0 classifier (74% accuracy)
  • Inspection pipeline with health scoring
  • PyTorch + ONNX + CoreML + TorchScript
  • 1,301 annotated anomaly images (3,071 bboxes)
  • Priority email support
Enterprise Solution
$2,999+
per project
  • Everything in Complete Pipeline
  • Custom model for your data
  • Web dashboard & REST API
  • Interactive inspection map
  • Edge deployment support (CoreML/TorchScript)
  • 60-day priority support

Resources

๐Ÿ”ง
Valve Detection Model
YOLOv8s, 92.95% mAP50, ONNX + PyTorch
โš ๏ธ
Anomaly Detection Model
YOLOv8s V13, 73.06% mAP50, 6 defect classes, ONNX + PyTorch + TorchScript
๐ŸŽฎ
Live Demo
Upload your inspection photo, see results
๐Ÿ“
Technical Blog
Full deep-dive into training methodology
๐Ÿ“ฆ
Dataset
9,038 images with 45K annotations
๐Ÿ’ป
GitHub
Source code, pipeline scripts, issues