E ² -NILM: Unified Representation Learning for Edge-Deployable High-Frequency Non-Intrusive Load Identification

Non-intrusive load monitoring (NILM) infers appliance states and switching events from aggregated electrical signals and is important for edge-side energy management and electrical safety sensing. High-frequency voltage and current waveforms provide rich transient and harmonic information, but their high dimensionality makes edge deployment difficult. Existing methods also insufficiently ex-ploit the complementary roles of transient, short-term steady-state, and frequency-domain features. This paper proposes E 2 -NILM, a lightweight high-frequency event-based NILM method for edge deployment. E 2 -NILM constructs event-centered multi-cycle windows, uses Gaussian soft labels to reduce sensitivity to annotation offsets, and extracts transient differential waveforms, short-term steady-state statistics, and low-order harmonic features. A three-branch encoder with gated fusion learns a shared representation for appliance category, operating-state, load-type, and aggregate-event recognition. A difficulty-aware early-exit mechanism further reduces average inference cost by rout-ing high-confidence samples to shallow exits and forwarding complex samples to the full fusion path. Experiments on TDHA25 show that E 2 -NILM achieves 0.972 Accuracy and 0.964 Macro-F1 for single-appliance recognition, and 0.926 socket-level Accuracy and 0.924 Micro-F1 for aggregate event recognition. Combining dynamic early exiting with knowledge distillation and structured prun-ing reduces average FLOPs from 19.4M to 11.8M and latency from 5.3 ms to 2.9 ms. With INT8 quantization, the energy-normalized recognition efficiency reaches 197.55 Acc/J. These results show that conditional multi-view feature fusion can improve the edge efficiency of high-frequency NILM while maintaining recognition performance.