Multifunctional system for the automatic inspection of flat broaching tools, featuring:
- 2-stage AI-based image processing system (detection of tooth application surface followed by failure segmentation)
- 3D analysis system with laser profilometer (measurement of tooth height and detection of surface defects)
- Module for cost optimisation of the tool reconditioning process (selection of a regeneration method that maximizes overall profit over the tool's life cycle)
- Dash based interactive application to view results, select the optimum regeneration method and generate a tool scan report
- SQL database for storing the results of the performed scans
- Detectron2 models based on the ResNET architecture
- Supervised learning using custom datasets
- Annotations with artifacts saved in .json format, compatible with LabelMe software
- System for semi-automatic datasets creation
- Jupiter notebooks with scripts for training and evaluation of AI models
- Proprietary evaluation methods with modified F1 metric
- Model comparison and performance tracking using Neptune.AI
- System for AI errors reporting (MLOps)
- Python 3.10
- Detectron2 v0.6
- PyTorch 1.12
- NVIDIA CUDA 11.7
Broach wear map.
Wear analysis for broach rows.
Preview of apposition surface and detected artifacts.
TKinter based host application with console output
MLOPs platform - Neptune AI for experiments tracking pourposes.
Automatic labels in .json lableme compatible format, generated using detectron2 model inference.