Multidimensional time-series perception data-driven fault feature extraction and recognition for power-plant auxiliary equipment

Abstract

Auxiliary equipment such as primary air fans, forced/induced draft fans and slurry circulating pumps are critical to the safe and continuous operation of large coal-fired generating units, yet they suffer frequent bearing, blade and lubrication faults whose early signatures are weak and easily masked by industrial noise. To address the limitation of singlemodality monitoring under conventional Distributed Control Systems, this paper proposes a multi-dimensional timeseries perception data-driven framework that jointly exploits vibration, acoustic, temperature and process variables for fault feature extraction and recognition. An integrated edge sensor node synchronously samples vibration up to 20 kHz, sound pressure at 48 kHz, surface temperature, and magnetic-field-derived rotational frequency, transmitting the streams over a two-wire Ethernet-APL link via MQTT. A Multi-Modal Cross-Attention Fusion (MM-CAF) network is then constructed, in which per-modality 1D-CNN-BiLSTM encoders capture local impulses and long-range temporal dependence, and a cross-attention fusion module weights complementary information across modalities. Validated on field data from a 1000 MW ultra-supercritical unit covering eight operating conditions, the proposed method achieves an overall recognition accuracy of 95.69%, outperforming vibration-only CNN and CNN-LSTM baselines by 10.2% and 6.8%, respectively, while sustaining sub-second edge inference, demonstrating its practicality for intelligent monitoring of power-plant auxiliaries.