Research Projects

Four independent, research-grade frameworks covering the core domains of industrial AI. Each repo is a self-contained platform with dataset loaders, model implementations, evaluation metrics, benchmark scripts, and tutorial notebooks.

Industrial Predictive Maintenance

GitHub: IEEE-IES-Industrial-AI-Lab/Industrial-Predictive-Maintenance

Industrial machines fail unexpectedly, causing costly unplanned downtime. This framework provides complete AI pipelines for Remaining Useful Life (RUL) prediction and fault detection on real industrial sensor data.

Models

Model Architecture Task Key Strength
LSTM 2-layer stacked LSTM + Bahdanau attention RUL regression Long-range temporal dependencies
Transformer Encoder-only + sinusoidal positional encoding RUL regression Parallel sequence modeling
TCN Dilated causal convolutions + residual connections RUL regression Exponential receptive field growth
Autoencoder LSTM encoder-decoder Anomaly detection Unsupervised — no fault labels required

Datasets

Dataset Domain Signals Fault Types
NASA CMAPSS Turbofan engine 21 sensors HPC/fan degradation
CWRU Bearing Rolling bearing Vibration (2ch) Outer/inner race, ball
IMS Bearing Rolling bearing Vibration (4ch) Outer/inner race, roller
Paderborn Bearing Rolling bearing Vibration + current 32 damage conditions

Key Features

  • Unified dataset loaders with automatic download for CMAPSS
  • Signal preprocessing: Butterworth filter, wavelet denoising, Z-score normalization
  • Time-domain features: RMS, kurtosis, crest factor, Shannon entropy
  • Frequency-domain features: FFT, PSD, spectral centroid, band energy
  • ONNX export + INT8 quantization for edge deployment
  • Real-time sliding window inference pipeline
  • 6 fully reproducible Jupyter notebooks

View Repo → Notebooks →

Industrial Time-Series AI

GitHub: IEEE-IES-Industrial-AI-Lab/Industrial-Time-Series-AI

Factories generate massive multivariate sensor streams that standard ML libraries handle poorly. This framework provides reproducible SOTA baseline implementations specifically designed for industrial IoT, ICS/SCADA systems, and smart-manufacturing data.

Models

Model Architecture Forecasting Anomaly Paper
LSTM Stacked RNN + linear head ✓ ✓ (AE) Hochreiter & Schmidhuber, 1997
TCN Causal dilated Conv1d + residual ✓ — Bai et al., 2018
Transformer Self-attention + positional encoding ✓ — Vaswani et al., 2017
PatchTST Patching + channel-independent attention ✓ — Nie et al., ICLR 2023
DLinear Trend/residual decomposition + linear ✓ — Zeng et al., AAAI 2023

Datasets

Dataset Domain Features Task Access
ETTh1/h2, ETTm1/m2 Power grid 7 Forecasting Free
PSM Server metrics 25 Anomaly detection Free
SMAP / MSL NASA telemetry Multi Anomaly detection Free
SWaT / WADI Water ICS 51 / 123 Both Request from iTrust

Key Features

  • Zero-copy strided window extraction for large sensor archives
  • Point-Adjust (PA) metric — standard for ICS/SCADA anomaly benchmarks
  • DLinear insight: strong linear baseline that matches Transformers on ETT
  • Built-in synthetic SWaT/WADI data — no download required to run benchmarks
  • 3 end-to-end tutorial notebooks

View Repo → Notebooks →

AI Power Electronics Diagnostics

GitHub: IEEE-IES-Industrial-AI-Lab/AI-Power-Electronics-Diagnostics

Industrial power electronics systems — inverters, motor drives, converters — fail due to switching faults, overheating, and harmonic distortions. This framework provides AI-based early fault detection from electrical signals (voltage waveforms, current signals, harmonic spectrum).

Models

Model Architecture Key Strength Parameters
1D CNN Residual Conv1d blocks + GAP Fast training, strong baseline ~1.2M
Spectrogram CNN ResNet-18 on STFT images Transient fault signatures ~11M
Transformer Patch-based encoder Long-range temporal patterns ~800K
BiLSTM + Attention 2-layer BiLSTM + additive attention Interpretable, sequential ~1.8M
Autoencoder 1D Conv encoder-decoder Unsupervised, no labels needed ~500K

Signal Processing Pipeline

Raw Signal (C, T)
       ├─→ FFT Analysis       → amplitude spectrum, THD, harmonic features
       ├─→ STFT Spectrogram   → (C, H, W) time-frequency image for CNN
       ├─→ Wavelet Features   → DWT sub-band energies, CWT scalogram
       └─→ Harmonic Analysis  → IEEE 519 compliance, ITSC sideband detection

Fault Domains

Domain Fault Classes Signals
3-Phase VSI Inverter 9 (IGBT T1–T6 open, short, DC fault, normal) Va, Vb, Vc, Ia, Ib, Ic
PMSM Motor Drive 5 (phase loss, ITSC, bearing, overtemp, normal) Ia, Ib, Ic

Key Features

  • Physics-informed synthetic data generators — no download needed
  • Unified SignalFeatureExtractor with FFT, STFT, wavelet, and harmonic modes
  • IEEE 519-2022 harmonic compliance checking built in
  • Streaming fault detection with SwitchFaultDetector and HarmonicFaultDetector
  • 4 tutorial notebooks from EDA to streaming inference

View Repo → Notebooks →

Smart Manufacturing AI

GitHub: IEEE-IES-Industrial-AI-Lab/Smart-Manufacturing-AI

Modern factories need AI across four interconnected challenges. This framework provides research-grade pipelines for visual defect detection, robot anomaly detection, RL-based scheduling, and digital twin simulation.

Modules

Module Problem Approach Key Models
vision/ Surface defect detection Fine-tuned + GradCAM ResNet-50, EfficientNet-B4, ViT-B/16
robotics/ Robot joint fault detection LSTM Autoencoder LSTM-AE variants
optimization/ Production scheduling Proximal Policy Optimization PPO (SB3)
digital_twin/ Production line simulation Discrete-event + MQTT/OPC-UA —

Datasets

Dataset Domain Size Access
MVTec AD Industrial surfaces & objects ~4.9 GB Research (non-commercial)
NEU Surface Defect Hot-rolled steel strip ~36 MB Free
Robot Sensor (synthetic) 6-DOF robot joints Generated locally Built-in

Key Features

  • Attention Rollout visualization for ViT-B/16
  • GradCAM heatmap overlay on defect images
  • ManufacturingEnv (Gymnasium-compatible) for RL training
  • TwinSimulator with TwinSync for MQTT/OPC-UA integration
  • 5 tutorial notebooks covering all modules

View Repo → Notebooks →