Online Equipment Fault Diagnosis

Mapping physical topology into hypergraph priors, combined with a physics-informed multi-task network and uncertainty-driven cascaded inference, achieves 99.55% compound-fault diagnosis accuracy

An operating-condition-aligned, risk-aware remaining-useful-life prediction framework that won first place worldwide, by a clear margin, at the PHM 2025 international data-challenge

A large-model agent diagnostic framework fusing rail images, time-series waveforms, and text logbooks achieves 96% visual rail-corrugation recognition accuracy, delivering fully automated fine-grained diagnosis under complex operating conditions

Production deployment of long-life-cycle diagnosis under varying operating conditions, running at General Motors and CITIC Dicastal

Acoustic-physical fusion for online detection and isolation of engine/bearing abnormal noise; patents granted in the US, Germany, and China

Servo motors are the joints of automated production lines, and their faults typically first surface as weak signatures in current and vibration signals. Following a "digital signal processing + expert rules + machine-learning enhancement" roadmap that starts from Motor Current Signature Analysis (MCSA), the team fuses fault-mechanism knowledge with strongly generalizing classifiers to deliver online motor monitoring and predictive maintenance, while evolving toward a next-generation paradigm of LLM-based retrieval-augmented diagnosis and representation learning. The technology now serves thousands of motors across hundreds of models, deployed at scale on production sites in papermaking, wind power, PCB manufacturing, precision forming, and other industries.

Elevators are among the most frequently used and safety-critical assets in high-rise buildings, yet fault samples such as shaking and emergency stops are scarce, and signal patterns vary widely across elevator models and service ages — thresholds and single-model approaches struggle to keep up. The team links car-mounted triaxial accelerometers, intelligent visual monitors, and a centralized monitoring platform into a live system, with an Attention-based Multi-Task Learning Bidirectional Long Short-Term Memory network (AT-MTL-BiLSTM) at its diagnostic core that automatically zeroes in on the critical moments around fault onset. Validated on operational data from real in-service elevators, the system achieves accuracy, recall, and F1 scores above 0.94 for all three fault types: front-back shaking, left-right shaking, and emergency stop.