Towards Intelligent Monitoring: Detecting Degraded Flame Torch Nozzles
summary
Uses structure-borne sound sensors mounted at the zone‑1 edge to avoid clipping, samples at 2 MHz sampling, and manual annotation yields 879 labeled patches fed through FFT-based features and simple classifiers. The pipeline consumes ~36 CPU‑days to explore windowing and downsampling; a feed‑forward net and SVC reach 0.97–0.99 balanced accuracy/F1 distinguishing degraded versus intact nozzles, but only one physical intact nozzle was available so generalization across nozzle sizes remains unverified.
AI-written summary · about
chapters (15)
- 0:00Degraded nozzle detection
- 1:11Flame cutting challenges
- 2:57Sensor selection experiments
- 4:42Danger zones and placement
- 7:55Test processes for nozzles
- 10:27Labeling data with Label Studio
- 13:39Feature extraction with FFT
- 15:24Train-test split and metrics
- 17:13Model training study
- 18:40Results and future outlook
- 21:52Q&A: Sensor details
- 22:57Q&A: Production changes
- 24:43Q&A: Model robustness
- 26:33Q&A: Production interest
- 27:57Q&A: Intact nozzle detection
related talks
38mHow We Built a Safety Violations Detection System
24mLeak Detection System using Machine Learning Techniques
43mDetecting Anomalies in Streaming Data
46mAI-Driven Techniques for Noise Filtration in Software Validation Logs
14mMinder: Faulty Machine Detection for Large-scale Distributed Model Training
08mDeep Transfer Learning for Computer Vision
26mOptimized Invariant Representation of Network Traffic for Detecting Unseen...
25mMachine Learning: Too Smart for its Own Good