Energy Overhead in Sensor Nodes: Embedded Machine Learning and Data Offloading via Low Power Wide Area Networks

Benjamin Karic; Nina Herrmann; Jan Stenkamp; Paula Scharf; Fabian Gieseke; Angela Schwering

cs.LGOct 28, 2025

Energy Overhead in Sensor Nodes: Embedded Machine Learning and Data Offloading via Low Power Wide Area Networks

Authors: Benjamin Karic, Nina Herrmann, Jan Stenkamp, Paula Scharf, Fabian Gieseke, Angela Schwering

Organizations: University of Münster · re:edu GmbH & Co. KG

Abstract

Image- and audio-based remote IoT applications offer a substantial opportunity to address ecological changes. However, to be effective, they must operate for long periods on limited power while also providing data via LPWAN. Previous studies have investigated the energy efficiency of isolated components such as radio transmission or data processing methods within an IoT system. However, a holistic evaluation is required to determine the main drivers of energy consumption in sensor nodes. To address this, we systematically analyze energy consumption across the entire sensor-node pipeline, from data acquisition to optional on-device CNN classification and wireless network transmission, using an ESP32-S3 microcontroller. We compare on-device CNN classification followed by result transmission with transmitting raw sensor data across three environmental monitoring scenarios. Thereby, we vary developer-controllable variables, notably the model architecture and the LPWAN technologies (LTE-M, NB-IoT, LoRaWAN). Our results indicate that on-device inference can reduce sensor node energy consumption by 1.7x to 6.9x compared to raw data transmission, depending on transmission frequency, with an acceptable loss in accuracy. Further, our results narrow down the energy tradeoff to variables IoT developers can control. Therefore, our findings provide empirical guidance for designing energy-efficient, long-range sensing systems and demonstrate that TinyML enables sustainable, autonomous deployments.

Figures & tables

Panel of Figure 1 — Figure 1: Examples of remote environmental monitoring applications using image and audio data and iot concepts. Both applications usually operate in locations without a static power supply and bad connectivity.

Figure 2: Energy consumption of an iot node per sensing cycle. Transmitting raw sensor data via LTE-M with MQTT (Cloud-ML) dominates the budget. Running CNN classification on-device, followed by transmitting the classification result (Embedded-ML), reduces the energy by the factor given above each bar. For image classification, an OV2640 camera module, SqueezeNet, and the plantvillage dataset are used; for audio classification, a digital MEMS microphone, tcresnet , and the Yellowhammer dataset (see Section 4.2 ) are used. Because LoRaWAN supports only small payloads, we could not measure the transmission of image and audio sensor data

Figure 4: Overview of the experimental design. Experiments and measurements report per-task metrics. Grey boxes are measurements; orange boxes are experiments where a set of independent variables is systematically varied.

Figure 5: Selected hardware for the case study: Seeed XIAO ESP32-S3 Sense with two interchangeable expansion boards, the Wio-SX1262 Kit with LoRa antenna and Sense expansion board, and Walter SoM v1.6 with Taoglas LTE antenna. One euro for scale.

	Seeed XIAO ESP32-S3 Sense & Wio-SX1262 Kit	Walter SoM v1.6
MCU	ESP32-S3R8	ESP32-S3R2
Memory	512KB + 8MB	512KB + 2MB
Peripherals	Camera, Mic, SX1262 LoRa Radio	Sequans GM02SP Modem
LTE-M	$\times$	✓
nbiot	$\times$	✓
LoRaWAN	✓	$\times$

Table 1: Microcontroller specifications

Figure 6: Overview of the model preparation pipeline for Embedded- ml deployment.

Parameter			Values
Optimizer			AdamW
Learning rate scheduling			Multi - step method
Image Classification	Batch size		$64$
	Transfer learning	Learning rates	$1\times 10^{-4},\,1\times 10^{-2}$
		weight_decay	$1\times 10^{-4},\,1\times 10^{-2}$
	Fine tuning	Learning rates	$1\times 10^{-5},\,1\times 10^{-3}$

Table 2: Hyperparameters used during training.

Figure 7: Depiction of the power profile measured with the Power Profiler Kit II. The selected part corresponds to one SqueezeNet plantvillage classification cycle, colorized by different application subtasks. Measured data was smoothed using a running window of size 50 for readability.

Data	Model	Type	Size (MB)	Top-1 (%)	Top-5 (%)
PV	MobileNetV2	Float	9.38	99.55	100.00
	MobileNetV2	Int	2.38	99.25	99.98
	SqueezeNet	Float	3.00	99.51	100.00
	SqueezeNet	Int	0.79	99.35	100.00
	MCUNet (in4)	Float	6.00	99.72	100.00
	MCUNet (in4)	Int	1.6	97.56	99.95

Table 3: Model sizes and accuracy metrics. The options with the preferable trade-off between accuracy and model size are emphasized.

Panel of Figure 8 — Figure 8: Average energy footprint of running a single CNN classification cycle across different model architectures and datasets on the test devices. A cycle includes model loading, data preprocessing, inference, and postprocessing.

Panel of Figure 9 — Figure 9: Energy footprint of transmitting class labels, images, and audio using different configurations of networks, protocols, and power-saving modes. Bold text labels were added for small bars that were not visible.

Data	LPWAN	Protocol / Spreading Factor	Wakeup & Connect ( $\mu$ Wh)	Payload ( $\mu$ Wh)	Release & Sleep ( $\mu$ Wh)	Total ( $\mu$ Wh)
Class	LTE-M	MQTT	4.34	1.58	414.90	420.83
	LTE-M	CoAP	4.73	3.33	406.04	414.10
	NB-IoT	MQTT	162.96	2.01	314.40	479.37
1 byte	NB-IoT	CoAP	160.51	3.13	310.45	474.10
1 byte	LoRaWAN	sf 7	6.45	8.65	20.12	35.23
	LoRaWAN	sf 10	5.96	35.70	24.30	65.95

Table 4: Energy in

\mu

Wh for the transmission task by different network settings in combination with eDRX. The options with the least energy consumption per data type are emphasized.

Data	Condition		Payload ( $\mu$ Wh)		Total ( $\mu$ Wh)
LoRaWAN (spreading factor)
	SF	Mean		SD	Mean	SD
Class	SF7	8.65		0.03	35.23	0.36
	SF10	35.70		0.08	65.95	0.26
	SF12	134.61		6.01	176.77	6.33
LTE-M (attenuation level)

Table 5: Sensitivity of transmission-related energy to varying coverage conditions. For cellular networks, the most efficient option per condition is highlighted in bold.

Panel of Figure 10 — Figure 10: Relative energy change at 30 dB attenuation compared with the 0 dB baseline. Bar fill encodes the radio technology, while hatching encodes the application protocol.

Figure 11: Average energy per cycle (application + 30 s sleep) for different paradigms and modes across three use cases, decomposed by sensor, classification, and transmission costs. Cloud-ML and Embedded-ML scenarios transmit via LTE-M/MQTT and LoRaWAN sf 10, respectively. Local inference for Embedded-ML is performed using the best energy-accuracy tradeoff models: TC-ResNet8 ( Audio Bird ), MobileNetV2 ( Image Bird ), and SqueezeNet ( Plant Disease ). Values above the bars represent the energy decrease factor compared to the Cloud-ML naive baseline.

Figure 12: Battery lifetime with different transmission cycles from 10 minutes to once a day. A 12.95 Wh battery with a self-discharge factor of 0.75% is assumed. In all scenarios, the Embedded-ML is the preferable option.

Internet of Things

The Internet of Things (IoT) connects networked sensors, devices, and systems for sensing, communication, and automated action.

See all 105 papers

Exploiting In-Sensor Computing for Energy-Efficient Earth Observation
May 31, 2026Luigi Capogrosso, Pietro Bonazzi, Loris Hoxhaj +1Remote Sensing Image Understanding Energy-Efficient ML
Interdisciplinary Transformation University of Austria · ETH Zurich · University of Verona
TinyML Enhances CubeSat Mission Capabilities
Mar 20, 2026Luigi Capogrosso, Michele MagnoOn-Device Inference Energy-Efficient ML
Interdisciplinary Transformation University of Austria · ETH Zurich
EdgeSpike: Spiking Neural Networks for Low-Power Autonomous Sensing in Edge IoT Architectures
Apr 29, 2026Gustav Olaf Yunus Laitinen-Fredriksson Lundstrom-Imanov, Taner YilmazEnergy-Efficient ML Edge Inference
Department of Economics, Stockholm University, Universitetsvägen 10 A, SE-106 91 Stockholm, Sweden · Department of Computer Engineering, Afyon Kocatepe University, 03200 Afyonkarahisar, Türkiye

May 31, 2026J/K move · Enter open · S save

Exploiting In-Sensor Computing for Energy-Efficient Earth Observation

Luigi Capogrosso, Pietro Bonazzi, Loris Hoxhaj, Michele Magno

Interdisciplinary Transformation University of Austria · ETH Zurich · University of Verona

Remote Sensing Image Understanding Energy-Efficient ML Remote Sensing In-Sensor Computing On-Device Inference

Open paper AskPDF arXiv

The rapid growth of the satellite industry has driven a significant increase in geospatial data acquisition, highlighting a critical bottleneck: the severe disparity between the volume of collected sensor data and the limited downlink bandwidth available to ground stations. While On-Board Computing (OBC) has helped address this by pre-processing data in orbit, this article further advances the paradigm by introducing an in-sensor computing framework. We present an optimized end-to-end Earth Observation (EO) pipeline tailored for strict computational constraints by integrating TinyML techniques with the Sony IMX500 Intelligent Vision Sensor. Specifically, our approach shifts processing directly to the sensor level, offloading the computation from the primary embedded device, and effectively mitigating the downlink transmission of noisy or irrelevant data. We evaluated several efficient Convolutional Neural Networks (ConvNets), i.e., SqueezeNet, ShuffleNetV2, and MCUNetV1, on the EuroSAT dataset. Experimental results show that, despite the optimizations required for deployment on the IMX500 platform, our models maintain a competitive 96.68% accuracy while operating within its 8 MB constraints. Specifically, the models reach an average processing throughput of 17.40 FPS with a latency of 27.43 ms. Furthermore, our system profile exhibits high energy efficiency, with a low energy footprint of 14.19 mJ per inference and an efficiency rating of 42.26 GMAC/J, demonstrating its viability for in-sensor deployment.

Abstract

Internet of Things

Explore similar work

Exploiting In-Sensor Computing for Energy-Efficient Earth Observation

TinyML Enhances CubeSat Mission Capabilities

EdgeSpike: Spiking Neural Networks for Low-Power Autonomous Sensing in Edge IoT Architectures