PerSANet: Synergizing Multi-Scale Inception and Energy-Based Attention for Time Series Anomaly Detection

Time series anomaly detection is a cornerstone of reliability in modern industrial systems and smart grids. While recent advancements, notably TimesNet, have achieved state-of-the-art performance by transforming 1D temporal sequences into 2D period-frequency tensors, they typically rely on single-scale convolutions. This design limitation struggles to simultaneously capture anomalies of varying durations—such as transient spikes and long-term trend shifts—resulting in a scale mismatch problem. To bridge this gap, PerSANet is introduced to harmonize multi-granularity feature extraction with energy-based saliency. Drawing inspiration from remote sensing, the Poly Kernel Inception module employs multi-scale depthwise convolutions to detect anomalies analogously to identifying small objects against complex backgrounds. SimAM, a parameter-free attention mechanism derived from neuroscience energy theories, is integrated to highlight high-energy anomalous regions without adding computational overhead. Experiments on real-world benchmarks show that PerSANet improves the precision–recall trade-off over strong baselines. Specifically, PerSANet achieves F1-scores of 96.47% on PSM, 91.02% on SWaT, and 84.99% on SMD, with improvements of 0.44–0.90 percentage points over TimesNet, consistently confirmed through multiple independent runs. The framework provides a practical, parameter-efficient upgrade for safety-critical industrial monitoring. The proposed approach is particularly valuable for applications requiring high detection reliability with minimal false alarms, such as water treatment systems, server monitoring, and smart grid fault detection.