Multidimensional perception data-driven alarm research and intelligent early-warning for power plant auxiliary equipment

Abstract

Boiler auxiliary equipment — primary air fans, forced draft fans, induced draft fans and slurry circulating pumps — is critical to the safe and economic operation of large coal-fired generating units. Conventional Distributed Control Systems (DCS) only acquire low-rate process variables such as bearing temperature, vibration RMS, blade angle and oil pressure, which is insufficient to capture incipient mechanical faults including bearing wear, blade jamming, oil-pressure fluctuation, motor overload, pump cavitation and unusual sound emission. To address this gap, this paper proposes a multi-dimensional perception data-driven alarm research and intelligent early-warning framework for the auxiliary equipment of a 1000 MW ultra-supercritical unit. An integrated vibration/acoustic/temperature/magnetometer sensor with synchronous high-rate sampling (vibration 0.1–20 kHz, voiceprint 48 kHz, surface temperature −40–120 °C) is developed and connected through Ethernet-APL and MQTT. At the edge, signal conditioning, time-frequency feature extraction and a lightweight 1-D convolutional model relieve cloud bandwidth. At the cloud, a cross-modal attention fusion network combines vibration, acoustic and process modalities, supported by a health-fingerprint library with more than 200 anomaly rules. The platform was deployed on Unit #2 of a 1000 MW coal-fired power plant; on a six-month field data set the proposed model achieves 95.7% accuracy and 95.2% F1-score across eight typical fault categories, raises the average alarm advance time by 38 minutes over single-modal baselines and reduces the false alarm rate to 2.1%.