Hardware-Aware Quantization and Pruning for Ultra-Low Power Environmental Anomaly Detection on Edge Nodes

Running artificial intelligence directly on low power sensor chips eliminates the need to transmit data to cloud servers, but standard microcontrollers feature less than 500 kilobytes of memory. While compression techniques enable edge classification, autoencoder anomaly detection requires precise decimal calculations to identify minor environmental deviations. This study evaluates an autoencoder subjected to structured pruning and strict 8 bit whole number quantization on environmental sensor telemetry. Pruning connections by up to 75 percent reduced memory consumption to 9.22 kilobytes, but forcing the network into an 8 bit format degraded the macro F1 score from 0.99 down to 0.4383. Quantization noise caused significant overlap between normal telemetry and fault conditions across the 99th percentile error threshold. These results indicate that strict integer formats obscure subtle anomaly signals, demonstrating that resource constrained edge nodes require decimal precision formats for reliable telemetry diagnostics.