Real-time Monitoring Technology for Power Safety Equipment Based on Self-supervised Online Hard Sample Mining and Multi-modal Sensor Fusion

Real-time monitoring for power aerial work safety faces challenges including high false alarm rates (15-20%) in posture monitoring due to single-sensor data deviations in complex electromagnetic environments, and difficulty in distinguishing loose from unlocked states in safety belt latch monitoring, alongside a lack of adaptive optimization. This study proposes a system integrating a 5G+LoRa multi-mode wireless sensor network, fusing accelerometer and angular rate sensor data for posture assessment via a Gaussian Mixture Model. Latch state is captured cooperatively by magnetic induction and mechanical sensing. The approach innovatively introduces a self-supervised online hard sample mining mechanism to optimize performance and constructs a multi-sensor fusion architecture to refine latch state logic. Experiments show the technology achieves a comprehensive performance index of 0.94 (a 27.8% increase), reduces posture monitoring false alarms to 3.2% (79.6% lower), and maintains a robustness coefficient of 0.78 under strong 120dBμV/m interference (25.8% higher than traditional methods). With a practicality index of 0.95, the technology enhances adaptability by 10.0%, providing accurate and efficient support for intelligent safety monitoring in power aerial work.