Industrial AI for
Bearing Fault Diagnosis
under Extreme Noise

Industrial AI Researcher | Physics-Informed Deep Learning | Predictive Maintenance & Fault Diagnosis

I am an AI/ML researcher specializing in physics-informed deep learning for industrial condition monitoring, predictive maintenance, and signal-based fault diagnosis. My work sits at the intersection of classical signal processing and modern neural architectures, with a focus on real-world reliability under harsh industrial conditions. My most recent research introduces a novel deep learning framework for bearing fault diagnosis that outperforms state-of-the-art baselines across five public benchmarks — including CWRU, IMS run-to-failure, Paderborn real fatigue data, and the NREL utility-scale wind turbine gearbox dataset — achieving up to 99.98% accuracy and remaining robust even at extreme −10 dB Signal-to-Noise Ratio conditions. The architecture is specifically engineered for edge deployment, with only 1.4M parameters and ≈8.5 ms inference latency. This work has been submitted to the International Journal of Machine Tools and Manufacture (Elsevier Q1). What I bring to your project: ✔ End-to-end deep learning pipeline: data → model → edge deployment ✔ Strong signal processing expertise (vibration, time-series, spectral analysis) ✔ Physics-informed model design that generalizes beyond the lab ✔ Rigorous, leakage-free evaluation with multi-trial statistical reporting ✔ Clean, reproducible PyTorch code with LaTeX-ready scientific documentation ✔ Hands-on experience with publicly recognized international benchmarks (CWRU, IMS, Paderborn, NREL)

Project: Physics-Informed Deep Learning for Industrial Bearing Fault Diagnosis Domain: Predictive Maintenance · Condition Monitoring · Wind Energy Status: Submitted to International Journal of Machine Tools and Manufacture (Elsevier Q1, 2025) Overview: Designed and implemented a novel deep learning framework (PhyQ-TransNet) for automated bearing fault detection in noisy, real-world industrial environments. The architecture combines a physics-informed nonlinear feature extractor with a Transformer encoder and a statistical uncertainty module — engineered to remain reliable at Signal-to-Noise Ratios where conventional CNN models fail. Key Achievements: • 99.98% accuracy on the IMS run-to-failure benchmark (real natural degradation data) • 99.03% accuracy on Paderborn real fatigue damage data (calibration ECE = 0.0047) • 99.72% accuracy at extreme −10 dB SNR on the NREL wind turbine gearbox dataset • Validated on 5 independent public benchmarks with a strict leakage-free chronological protocol • Edge-deployable: 1.40M parameters · 3.21 GFLOPs · ~8.5 ms inference latency (~117 FPS) Methodology highlights: • Physics-informed design: the model is structured to mirror physical signal demodulation — no black-box approach • Dual-stream architecture fusing deep feature learning with robust statistical priors • Noise resilience: performance validated under Gaussian noise 10× stronger than the fault signal • Multi-sensor fusion: synchronous data from three accelerometers per experiment Tech Stack: Python · PyTorch · Signal Processing · Transformer Encoder · Physics-Informed ML Affiliation: Mechatronics Engineering, Urmia University of Technology, Iran

I have hands-on research experience building and validating deep learning systems for industrial fault diagnosis under real-world noise conditions. My recent work — currently under review at an Elsevier Q1 journal — demonstrates state-of-the-art performance on five internationally recognized bearing and gearbox benchmarks, including the NREL wind turbine dataset at −10 dB SNR. The resulting model runs in under 9 ms per inference and fits within 5.4 MB of memory, making it suitable for edge deployment on industrial controllers. I approach every project with the same rigor: physics-informed design, leakage-free evaluation, and reproducible code.