Intelligent Anomaly Detection Method for Consumer Electronics Based on Feature-Driven Learning and Diffusion Model

The increasing complexity of consumer electronics and their dependence on continuous multivariate sensor data make intelligent anomaly detection a critical task. This paper proposes a novel anomaly detection framework, termed FD-Diffusion (Feature-Driven Diffusion Framework), tailored to time-series data generated by consumer electronic devices. The framework is composed of three key components. First, the Status-Aware Multimodal Encoder (SAME) extracts temporally aligned and semantically enriched representations from heterogeneous sensor streams by incorporating device status, behavioral context, and operational states. Second, the Condition-Aware Diffusion Model with Residual Refinement (CADM) leverages these embeddings to guide the reverse denoising trajectory of a diffusion process, ensuring that reconstructed sequences conform to normal behavior patterns. To enhance sensitivity to subtle deviations, CADM includes a residual refinement decoder with localized attention. Third, the Multi-Scale Adaptive Anomaly Scoring Mechanism (MAASM) fuses reconstruction loss, latent divergence, and semantic inconsistency into an interpretable and adaptive anomaly score, adjusted dynamically based on recent operational statistics. Experimental results on real-world device datasets demonstrate that FD-Diffusion outperforms existing methods in both detection accuracy and interpretability, offering a scalable solution for intelligent monitoring of consumer electronics under dynamic and diverse usage conditions.