Edge Intelligence for Low-Power IoT Sensing: Balancing Accuracy and Energy through Model Compression

The proliferation of Internet of Things (IoT) devices necessitates
processing paradigms that move beyond traditional cloud-centric
models, particularly for low-power embedded sensing where
connectivity, latency, and energy budgets are tightly constrained. This

paper presents an extended investigation of edge intelligence for low-
power IoT sensing, focusing on the integration of machine learning

(ML) directly on resource-constrained devices and the resulting trade-
off between computational accuracy and energy consumption. We

evaluate five edge-deployable model configurations a Baseline CNN,
ResNet-50, MobileNetV2 (INT8), a pruned SqueezeNet, and a
Proposed Method combining quantization, structured pruning, and a
depth-wise separable architecture on the CIFAR-10 benchmark, and
profile each model's per-inference energy consumption on a
representative Cortex-M7 microcontroller. The Proposed Method
achieved the highest accuracy (0.934) and F1-score (0.928) while
consuming only 15.7 mJ per inference, 8.4 times less energy than
ResNet-50 despite a smaller accuracy gap of 4.3 percentage points. A
bit-width sweep further shows that INT8 quantization preserves
accuracy within 1.3 percentage points of full-precision inference while
cutting energy consumption by roughly 88%, whereas aggressive 2-bit
quantization causes a sharp accuracy collapse. These findings quantify
the accuracy-energy Pareto frontier for on-device inference and
demonstrate that carefully combined compression techniques — rather
than any single technique in isolation — offer the most favorable
operating point for battery- and energy-harvesting-powered IoT
sensors. We conclude with design guidance for selecting compression
strategies according to application-specific accuracy and energy

constraints, and outline directions for future work on adaptive, runtime-
configurable inference.