Lightweight Deep Learning for Real-Time Anomaly Detection on Intelligent Edge Devices

The proliferation of Internet-of-Things sensors has created a need to detect anomalies where data are generated, yet deep anomaly detectors are typically too large and power-hungry for microcontroller-class devices at the network edge. Streaming raw data to the cloud incurs latency, bandwidth costs, and privacy risks, so on-device inference is increasingly preferred, but it demands models that fit within kilobytes of memory and run within tight energy budgets. This paper presents LiteAD, a lightweight deep learning framework for real-time anomaly detection on intelligent edge devices. The approach combines a depthwise-separable convolutional autoencoder trained only on normal data with structured pruning and integer quantization, scoring anomalies by reconstruction error against an adaptive threshold. The framework is described formally and as pseudocode, covering offline compression and on-device inference. Using an illustrative synthetic time-series benchmark, LiteAD is compared against four baselines on detection quality, latency, memory, and energy. The synthetic results indicate that the compressed model retains detection accuracy close to a full-precision autoencoder while reducing model size and inference latency by roughly an order of magnitude. These findings support lightweight deep learning as a practical approach to always-on anomaly detection at the edge.