Content area
Industrial pump systems require real-time fault diagnosis for predictive maintenance, but conventional Chain-of-Thought (COT) reasoning faces computational bottlenecks when processing high-frequency vibration data. This paper proposes Vibration-Aware COT (VA-COT), a novel framework that integrates multi-domain feature fusion (time, frequency, time–frequency) with adaptive reasoning depth control. Key innovations involve expert prior-guided dynamic feature selection to optimize edge-device inputs, complexity-aware reasoning chains reducing computational steps by 40–65% through confidence-based early termination, and lightweight deployment on industrial ARM-based single-board computers (SBCs). Evaluated on a 12-class pump fault dataset (5400 samples from centrifugal/gear pumps), VA-COT achieves 93.2% accuracy surpassing standard COT (89.3%) and CNN–LSTM (Convolutional Neural Network-Long Short-Term Memory network) (91.2%), while cutting latency to <1.1 s and memory usage by 65%. Six-month validation at pump manufacturing facilities demonstrated 35% maintenance cost reduction and 98% faster diagnostics versus manual methods, proving its viability for IIoT (Industrial Internet of Things) deployment.
Details
Deep learning;
Wavelet transforms;
Artificial neural networks;
Real time;
Signal processing;
Computer applications;
Gear pumps;
Long short-term memory;
Vibration analysis;
Machine learning;
Fault diagnosis;
Maintenance costs;
Fourier transforms;
Neural networks;
Process controls;
Reasoning;
Support vector machines;
Network latency;
Industrial applications;
Computers;
Vibration;
Latency;
Predictive maintenance;
Industrial Internet of Things
1 College of Mechanical and Electrical Engineering, Wenzhou University, Wenzhou 325035, China; [email protected] (J.Z.); [email protected] (Z.L.)
2 Zhejiang TONGLI Transmission Technology Co., Ltd., Ruian 325205, China