We convert heterogeneous engineering measurements into physically meaningful fault evidence, mechanism-consistent explanations, and credible health decisions under variable operating conditions.
This direction treats diagnosis as an evidence-construction problem rather than a closed-set classification task. Signal processing, physical models, causal structures, and uncertainty descriptions are designed together so that a conclusion can be traced back to measurable phenomena.
Research covers rotating machinery, multistage transmissions, manufacturing processes, and other condition-monitoring systems in which faults evolve across components, operating regimes, and sensing modalities.
Speed, load, sensor position, noise, and structural transfer paths alter the observed waveform. The challenge is to retain fault order, modulation, recurrence, propagation, and degradation information while separating operating-condition variation from mechanism-related structure.
Reliable diagnosis must explain how evidence propagates through components and subsystems. We combine physical priors, digital twins, event and causal graphs, transfer-path analysis, and temporal logic to distinguish correlation from mechanism-consistent explanation.
Models must remain useful with limited labels, unseen domains, compound faults, and uncertain evidence. We study calibrated uncertainty, cross-domain invariance, verifiable explanations, and decision interfaces for maintenance, isolation, and fault-tolerant operation.
Explore our research methods and digital-twin software in a 20-slide presentation, with a Chinese-subtitled demonstration on slide 19.
网页演示保留PPT实际渲染的图表、排版与渐变；原始PPT中的文字、图表和框架仍可编辑。视频为离线软件演示，不代表现场执行。
Vibration, current, force, images, operational logs, and physical models
Orders, time-frequency structures, events, causal graphs, and transfer paths
Domain generalisation, uncertainty reasoning, logic rules, and causal attribution
Fault identification, root-cause isolation, maintenance planning, and risk-aware action
