Enterprise AI Analysis

Model Selection & Performance on Edge Devices

We evaluated MobileNetV2 and SqueezeNet, fine-tuning them on PlantVillage (plant disease) and CUB (bird species) datasets. Crucially, models underwent 8-bit post-training quantization to reduce memory footprint, a necessity for resource-constrained microcontrollers like the ESP32-S3.

Quantization led to minor accuracy losses (around 1% for PlantVillage, 3-6% for CUB) while significantly reducing model sizes. For instance, quantized SqueezeNet models were approximately 8% the size of unquantized MobileNetV2 models.

Inference consistently accounts for over 90% of the total ML-related energy, highlighting its intensive nature despite model compression. This contradicts simple FLOPs-based estimations, emphasizing the need for real-world energy measurements on target hardware.

Model Performance & Size Comparison (Quantized 8-bit Integer)

Optimizing Data Transfer with LPWANs

Data transmission is often the most energy-intensive part of IoT operations. We benchmarked various LPWAN technologies and application layer protocols:

• LoRaWAN: Ideal for transmitting small data (e.g., class indices) due to low connection energy, but with duty cycle limitations for frequency.
• LTE-M & NB-IoT (MQTT/CoAP): Suitable for larger data, with LTE-M MQTT proving most efficient for transmitting full images. NB-IoT generally incurs higher latencies and energy costs.

Crucially, reducing message size from a full image to a small class index enables the use of highly energy-efficient protocols like LoRaWAN, which would otherwise be unfeasible.

Energy Footprint: Transmitting Class Label vs. Image

Sending a raw 224x224 pixel image via NB-IoT MQTT costs nearly 1.25 mAh, while sending a class index via LoRaWAN costs only 50.54 µAh. This stark difference highlights the immense energy savings from on-device inference.

Achieving Maximum Energy Efficiency in Practice

By combining on-device CNN inference with optimized data transmission, overall energy consumption can be drastically reduced. Our experiments show that executing CNN inference on-device and transmitting only the results reduces energy consumption by a factor of up to five compared to sending raw image data.

Furthermore, implementing a "result-based sending" strategy, where data is only transmitted when an event of interest is detected (e.g., every 10th measurement), leads to an energy reduction factor of almost 14 times.

These findings empower developers to create IoT applications with extended operational lifespans for remote environmental monitoring, significantly lowering the carbon footprint and maintenance costs.

Considerations & Future Directions

While EmbeddedML offers significant advantages, it comes with certain trade-offs:

• Loss of Raw Data: On-device processing means raw images are not stored or transmitted, limiting post-hoc analysis or re-evaluation.
• Model Constraints: The current approach is limited to relatively small models compressed for specific microcontrollers. Larger or more complex models might require more powerful (and energy-intensive) hardware.
• Hardware Specificity: The use of manufacturer-provided libraries (like ESP-DL) optimizes for specific hardware but reduces cross-platform compatibility.

Future work could explore Neural Architecture Search (NAS) to design energy-optimized models from scratch, investigate image batching for more efficient transmission of larger images, and evaluate protocol-specific energy-saving features like faster reconnection times.