Lightweight Feature-Focused FCNN for Power Equipment Health Monitoring on Edge Devices

Industrial power Internet of Things (IIoT) enables real-time monitoring of critical equipment like transformers but faces deployment challenges on resource-limited edge devices. Traditional threshold methods exhibit limitations in detection fidelity, while complex deep models incur prohibitive computational latency. To resolve this, we propose a Lightweight Fully Connected Neural Network (LFCNN) optimized for edge deployment. The model employs a feature-centric three-layer architecture prioritizing dominant fault indicators like current mutations and integrates a dynamic training algorithm using a two-stage learning rate scheduler with gradient clipping to accelerate convergence. Evaluated on operational substation data, the model demonstrates efficient real-time detection capability for abrupt anomalies, achieving high accuracy and low inference latency with modest memory requirements. Results confirm the effectiveness of the feature-centric design and training optimizations for edge compatibility. Limitations exist in subtle scenarios involving coupled temperature-vibration deviations near normal thresholds, indicating a direction for future multimodal fusion research.