We build deep-learning systems that detect, classify and forecast faults in rotating machinery — from distributed bearing damage to inverter failure — using vibration, stray magnetic flux and motor current. No new instrumentation on the shaft, no downtime to measure.
About the centre
The Machine Health Intelligence Research Centre works on condition monitoring, fault diagnosis and prognostics for electrical machines. Our focus is the hard case rather than the textbook one — distributed bearing damage that produces no clean defect frequency, compound faults that mix mechanical and electrical signatures, and models that must hold up when the load, the speed or the measurement session changes.
Our datasets come from bearings degraded by lubrication loss, contamination, electrical erosion and flaking — the failure modes that actually dominate industry, recorded on instrumented testbeds rather than synthesised by notching a race.
Three-axis vibration, stray magnetic flux and phase currents each see a different part of the failure. We fuse them in architectures designed for the fusion itself, and we measure what each channel is actually worth.
Cross-load, cross-speed and leave-one-session-out evaluation, with significance testing over many fits. We report where a method fails as carefully as where it works, because a diagnosis that only holds in one session is not a diagnosis.
Research areas
Six active lines, from sensing hardware through to deployment-grade generalisation — plus a second stream applying the same signal-intelligence methods to biomedical diagnostics.
Distributed damage spreads energy across the spectrum instead of concentrating it at a defect frequency, which is why classical envelope analysis struggles. We design 1-D, 2-D and 3-D convolutional architectures with hybrid inputs — raw time frames alongside engineered features — that separate lubrication loss, contamination, erosion and flaking from a healthy baseline across the full speed–torque envelope.
Fitting an accelerometer to a small cooling-fan motor is rarely practical or affordable. We estimate remaining lifespan from three-phase current alone, using residual and channel-attention blocks — in low-rated-torque machines the modulated torque disturbance from a degrading bearing is clearly legible in the phase currents.
A custom sensory board integrating three-axis vibration, a fluxgate stray-flux sensor and two-phase current acquisition — plus the study of how far ADC resolution and sampling rate can be reduced before diagnostic accuracy degrades.
A model that only works at the speed it was trained on has no industrial value. We use multimodal contrastive learning and order-domain representations to build diagnoses that survive changes in speed, load and acquisition session.
Beyond bearings: open-circuit and switch faults in PMSM drive inverters, broken-bar signatures, eccentricity, and the compound case where mechanical and electrical faults are present at once and their signatures interfere.
A bearing race and a heart are both systems whose condition is encoded in a noisy periodic signal. The same methods carry over: ECG arrhythmia classification, Parkinsonian gait analysis, sleep apnoea detection and musculoskeletal imaging with explainable attribution.
Capabilities
We work with industry and with academic groups. Engagements typically start with a feasibility study on your own data and end with a model you can run on your own hardware.
You send representative recordings from your machines. We report which faults are separable from the signals you already have, which extra channel would buy the most, and what accuracy is realistic before anyone commits to a rollout.
End-to-end: signal conditioning, feature and representation design, architecture selection, and an evaluation protocol fixed in advance so the reported number is the one you will actually see in the field.
Minimal feature sets and reduced ADC resolution can cut the cost of a monitoring node substantially. We quantify exactly how much accuracy each reduction costs, so the trade-off is a decision rather than a guess.
Remaining-useful-life models with calibrated uncertainty, so a maintenance window can be planned against a probability rather than a point estimate.
An outside read on an existing diagnostic pipeline: leakage between train and test, session confounds, over-optimistic protocols, and whether the reported accuracy would survive a change of load.
Co-supervision of graduate work, collaboration on funded projects and grant applications, and short technical courses on signal processing and deep learning for machine health.
Approach
The same five steps, whether the subject is a 10-hp induction motor or a twelve-lead ECG.
Vibration resampled on shaft angle fixes bearing and eccentricity coordinates. Stator-current signatures are anchored to supply frequency and slip instead, so forcing both onto one axis smears one of them. Getting the representation right is worth more than any architecture choice downstream.
Data across ten speed levels and five torque levels, healthy and faulted, so the model sees the variation it will meet in service rather than a single convenient operating point.
Vibration, flux and current are combined in architectures built for multi-input fusion — then ablated, so we know what each modality contributes rather than assuming more sensors is better.
Change the load. Change the speed. Hold out an entire acquisition session. The gap between those numbers and the in-distribution number is the honest measure of whether a model is deployable.
Minimal feature subsets, lower sampling rates, fewer ADC bits, smaller models. A diagnosis that needs a laboratory to run is a result; one that runs on a monitoring node is a product.
Selected publications
Published in IEEE Transactions, IEEE journals and IEEE conference proceedings.
Team
Founder & Director
B.Sc. in Biomedical Engineering (Sahand University of Technology), M.Sc. in Control Systems Engineering (University of Tehran), Ph.D. in Electrical and Electronic Engineering (Van Yüzüncü Yıl University, 2025). Works on evolutionary computation and optimisation, deep learning, computer vision and large language models, applied to machine health and biomedical diagnostics.
Co-founder
B.Sc., M.Sc. and Ph.D. in Electrical and Electronics Engineering from Van Yüzüncü Yıl University, where he is now an Assistant Professor. He has published widely and worked on research projects developing efficient energy conversion systems. His interests are power electronics, electrical machines, renewable energy systems and the control of electric drives.
Across three universities
Our published work is joint with researchers at the University of Texas at Dallas, Texas A&M University–Commerce and Van Yüzüncü Yıl University. We are open to new collaborations on condition monitoring, prognostics and applied signal intelligence.
Contact
If you have machines you would like to monitor, data you would like assessed, or a research collaboration in mind, write to us. A short description of your setup and what you are trying to detect is enough to begin.