Skip to content

How Does Hawk Use W5500 Ethernet for Low-Power Appliance Recognition?

Hawk is a SenSys 2024 ML system that uses ESP32/W5500 waveform acquisition and FFT-XGBoost to recognize low-power appliance events.

Benjamin

Published June 26, 2026

Original author: Zijian Wang, Xingzhou Zhang, Yifan Wang, Xiaohui Peng, Zhiwei Xu ยท ChinaOriginal source (new tab)

How Does Hawk Use W5500 Ethernet for Low-Power Appliance Recognition?

Components

Hardware components

WIZnet parts

Project description

๐Ÿ“Œ Overview

Hawk is a machine-learning system that recognizes when low-power appliances turn ON or OFF by observing the current waveform at a building's main power line. Its data-acquisition board combines a 24-bit ADS131M08 ADC, a dual-core ESP32, and W5500 Ethernet. The recognition pipeline then uses SsDiff signal processing, FFT features, and an XGBoost classifier to identify appliance events.

Hawk was published as an academic research paper. Researchers from the Institute of Computing Technology, Chinese Academy of Sciences, the University of Chinese Academy of Sciences, and Great Bay University presented it at ACM SenSys 2024. SenSys is the ACM Conference on Embedded Networked Sensor Systems, an international research venue covering sensors, embedded systems, networking, and IoT. The Hawk paper appeared in its Applications and Deployments program and received the conference's Best Artifact Evaluation Award.

Hawk in one line: ADS131M08 measures the waveform, ESP32 organizes and synchronizes the samples, W5500 sends acquisition data over Ethernet, and FFT-XGBoost recognizes the appliance event.

Hawk laboratory acquisition board and deployed electrical-panel prototype Source: Wang et al., Hawk, Figure 8, arXiv:2410.16293v1, CC BY 4.0.

What Is NALM, and Why Are Low-Power Appliances Difficult?

NALM stands for Non-Intrusive Appliance Load Monitoring. During recognition, instead of attaching a permanent sensor to every appliance, a NALM system measures the combined electrical load at one main-meter point and estimates which appliance changed state. It is similar to listening to several instruments playing together and identifying which one has just started or stopped. Hawk temporarily measures individual appliance currents while building labeled training data, then uses the aggregate main-meter stream for deployed recognition.

Low-power devices are difficult because their changes can be small compared with refrigerators, air conditioners, and other background loads. Hawk therefore looks beyond a single average-power value. It captures detailed current waveforms and learns the electrical pattern associated with each target appliance's ON and OFF events.

The project has two connected stages. First, it builds a balanced, labeled waveform dataset called HawkDATA. Second, it trains and evaluates a model that recognizes appliance events from the aggregate main-meter current.

Inside the ESP32 and W5500 Acquisition Board

The ADS131M08 can sample up to eight analog channels simultaneously at 16 kHz with 24-bit resolution. In a 50 Hz power system, one AC cycle lasts 20 ms, so Hawk records 320 waveform points per voltage cycle. This detail helps expose harmonic patterns that may be difficult to see in average power readings.

The ESP32 divides the acquisition work between its two cores. Core1 responds to ADC interrupts, reads samples, and packages them by voltage cycle. Core0 checks packet status and moves each completed cycle toward the logging path. The board then uses its two network interfaces for different tasks:

  • ESP32 Wi-Fi reads the router's TSF time information. Hawk uses the AC voltage cycle observed by every board as shared perceptible time, while TSF, or Time Synchronization Function, helps identify the cycle ID.
  • W5500 Ethernet carries the packaged measurement data through the router to the logging PC during HawkDATA construction. The system uses MQTT between the acquisition environment and the host.

Hawk acquisition architecture with Wi-Fi timing and W5500 Ethernet roles Source: Wang et al., Hawk, Figure 7, arXiv:2410.16293v1, CC BY 4.0.

Figure 7 shows the role split directly: Wi-Fi participates in the timing path, while W5500 provides the wired measurement-data path. This lets the ESP32 concentrate on sampling and cycle packaging while the logger receives the waveform records used to build the dataset.

From Labeled Waveforms to HawkDATA

Machine learning needs examples in which the waveform and the correct appliance event are paired. Hawk creates those labels with a balanced Gray-code schedule. Only one target appliance changes state at each step, producing a clear event such as "humidifier ON" or "monitor OFF." Randomized appliance groups and starting states expose the model to many background-load combinations.

The laboratory setup used 22 appliances: 18 programmable target appliances and four background devices. The full raw dataset contains 32.2 hours of measurements, 5,796,650 voltage cycles, 4,558 unique appliance-state combinations, and 127.8 GB of data.

Hawk also introduces shared perceptible time to match measurements from different nodes. Every node can observe the same AC voltage cycle, and the TSF value helps identify that cycle consistently. The paper reports a maximum synchronization error of 61.80 microseconds, which is below the 62.5-microsecond interval between samples at 16 kHz.

CDF of Hawk sampling synchronization error compared with TSF Source: Wang et al., Hawk, Figure 10, arXiv:2410.16293v1, CC BY 4.0.

How SsDiff, FFT, and XGBoost Recognize an Event

Hawk's selected recognition path can be read as four steps:

  1. SsDiff highlights the change. Steady-state Differential current subtracts a waveform from one captured 30 AC cycles earlier. Loads that remain unchanged largely cancel out, while the device that switched leaves a clearer difference.
  2. FFT creates a compact electrical fingerprint. FFT means Fast Fourier Transform. It converts a waveform changing over time into its frequency components. Hawk keeps the real part, imaginary part, and magnitude of the first ten harmonics, reducing a 320-point waveform to 30 features.
  3. XGBoost classifies the fingerprint. XGBoost is a gradient-boosting model that builds decision trees in sequence, with later trees learning to correct earlier errors. The classifier selects among 37 states: ON and OFF for 18 target appliances, plus IDLE.
  4. A 30-cycle vote confirms the event. Hawk checks whether the same prediction appears often enough across a short window, reducing isolated false detections.

Hawk SsDiff, FFT-XGBoost classification, and vote-based inference pipeline Source: Wang et al., Hawk, Figure 6, arXiv:2410.16293v1, CC BY 4.0.

Training and inference happen in different places. The model learns from labeled HawkDATA offline. In the field system, a C inference program runs on a Jetson Orin NX 16GB and subscribes to the aggregate-current stream over MQTT. The paper reports 4.07 ms average inference latency and 40.19 MB memory usage on that platform. The ESP32 and W5500 are the collection and communication front end; the Jetson runs the trained recognition model.

Results in the Dataset and in Real Spaces

Hawk reports F1 score, a metric that balances precision and recall. In plain language, it rewards a system for finding real appliance events while avoiding false reports. The paper presents each result for a specific evaluation:

  • 92.71% average event-recognition F1 on HawkDATA
  • about 93.94% state-recognition F1 on HawkDATA
  • 97.07% event-classification F1 on the separate BLUED dataset
  • 96.02% and 94.76% event F1 in residential and office deployments, respectively

The field study covered a 103 mยฒ residence and a 25 mยฒ graduate-student office with unknown background appliances. The floor plans below show where the main-meter sensors and evaluated appliances were placed.

Hawk residential and office deployment floor plans Source: Wang et al., Hawk, Figure 9, arXiv:2410.16293v1, CC BY 4.0.

โš™๏ธ Why the W5500 Role Matters

Hawk is a useful example of W5500 supporting an ML system without being the ML processor. During dataset construction, the Ethernet controller moves cycle-packaged, high-resolution measurements from the ESP32 acquisition board to the logging PC. Those records become the labeled data used to develop and evaluate the appliance-recognition model.

The component roles form a clear chain:

  • ADS131M08 converts analog current and voltage signals into synchronized digital samples.
  • ESP32 handles acquisition, voltage-cycle packaging, and timing queries.
  • W5500 supplies the wired Ethernet path for the measurement data.
  • The logging environment stores and organizes HawkDATA.
  • Jetson Orin NX executes real-time inference from the main-meter stream.

For WIZnet developers, the reusable design idea is the boundary between embedded sensing and machine learning. A network controller can connect a high-rate acquisition device to the host where datasets are built and models are run.

Public Research Artifacts

The authors provide the Hawk paper, the SenSys24-Hawk artifact repository, and the HawkDATA evaluation package. The repository includes evaluation scripts, XGBoost models, BLUED event data, synchronization analysis, and a short field video. SenSys 2024 recognized this research package with its Best Artifact Evaluation Award.

These materials are well suited to studying the dataset, synchronization method, model evaluation, and deployment results. The embedded acquisition board would require additional engineering work because its firmware, schematic, PCB, BOM, and broker configuration are not part of the artifact release.

Related WIZnet Maker Reading Path

For another multi-channel acquisition system that hands data to an AI workflow, see How Does W5500 Stream 8-Channel RF Data for Drone Detection?. It uses STM32 and AD7606 for RF sensing, while Hawk uses ESP32 and ADS131M08 for power waveforms.

TheBreeze: W5500 Ethernet DDP Receiver for WLED shows a similar wired/wireless role split in an LED system: W5500 handles the main pixel stream while Wi-Fi supports lighter control.

For a more practical energy gateway, ESP32 JSY Energy Meter Gateway with W5500 Support sends processed meter telemetry rather than Hawk's high-resolution training waveforms. Together, the two projects show the difference between an installable energy gateway and an academic ML data pipeline.

Source-Backed Summary

Hawk connects embedded measurement to appliance-recognition ML in a traceable sequence: capture with ADS131M08, organize and synchronize with ESP32, transport with W5500, extract harmonic features with FFT, classify with XGBoost, and confirm with 30-cycle voting. Its paper, dataset, award-winning artifact repository, and field evaluation make it a detailed academic reference for Ethernet-connected AIoT data acquisition.

โ“ FAQ

Q. What does NALM mean? NALM is Non-Intrusive Appliance Load Monitoring. It estimates individual appliance activity from an aggregate main-meter measurement instead of placing a sensor on every appliance.

Q. What do FFT and XGBoost each do? FFT converts the current waveform into harmonic features. XGBoost uses those features to classify which appliance switched ON or OFF.

Q. What are the separate roles of Wi-Fi and W5500? ESP32 Wi-Fi reads TSF timing information for synchronization. W5500 sends packaged acquisition data over wired Ethernet to the logging PC during dataset construction.

Q. Where do training and real-time inference run? Training uses labeled HawkDATA offline. The deployed C inference program runs on a Jetson Orin NX 16GB and reads the main-meter stream over MQTT.

Q. Can the complete acquisition board be rebuilt from the artifact repository? The repository supports study of the algorithms and evaluations. A full hardware rebuild also needs acquisition firmware and board-design details that are not included in the release.

ํ•œ๊ตญ์–ด (Korean)

๐Ÿ“Œ ๊ฐœ์š”

Hawk(ํ˜ธํฌ)๋Š” ๊ฑด๋ฌผ์˜ ๋ฉ”์ธ ์ „๋ ฅ์„ ์—์„œ ์ธก์ •ํ•œ ์ „๋ฅ˜ ํŒŒํ˜•์„ ๋ณด๊ณ  ์ €์ „๋ ฅ ๊ฐ€์ „์ด ์–ธ์ œ ์ผœ์ง€๊ณ  ๊บผ์กŒ๋Š”์ง€ ์ธ์‹ํ•˜๋Š” ๋จธ์‹ ๋Ÿฌ๋‹ ์‹œ์Šคํ…œ์ž…๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ์ˆ˜์ง‘ ๋ณด๋“œ๋Š” 24-bit ADS131M08 ADC, ๋“€์–ผ์ฝ”์–ด ESP32, W5500 Ethernet์œผ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. ์ดํ›„ SsDiff ์‹ ํ˜ธ ์ฒ˜๋ฆฌ, FFT ํŠน์ง• ์ถ”์ถœ, XGBoost ๋ถ„๋ฅ˜๋ฅผ ๊ฑฐ์ณ ๊ฐ€์ „์˜ ON/OFF ์ด๋ฒคํŠธ๋ฅผ ์ฐพ์•„๋ƒ…๋‹ˆ๋‹ค.

Hawk๋Š” ํ•™์ˆ  ์—ฐ๊ตฌ ๋…ผ๋ฌธ์œผ๋กœ ๋ฐœํ‘œ๋œ ์‹œ์Šคํ…œ์ž…๋‹ˆ๋‹ค. ์ค‘๊ตญ๊ณผํ•™์› ์ปดํ“จํŒ…๊ธฐ์ˆ ์—ฐ๊ตฌ์†Œ, ์ค‘๊ตญ๊ณผํ•™์›๋Œ€ํ•™, Great Bay University ์—ฐ๊ตฌ์ง„์ด ๊ฐœ๋ฐœํ•ด ACM SenSys 2024์—์„œ ๋ฐœํ‘œํ–ˆ์Šต๋‹ˆ๋‹ค. SenSys๋Š” ACM Conference on Embedded Networked Sensor Systems์˜ ์•ฝ์ž๋กœ, ์„ผ์„œ, ์ž„๋ฒ ๋””๋“œ ์‹œ์Šคํ…œ, ๋„คํŠธ์›Œํฌ, IoT ์—ฐ๊ตฌ๋ฅผ ๋‹ค๋ฃจ๋Š” ๊ตญ์ œ ํ•™์ˆ ๋Œ€ํšŒ์ž…๋‹ˆ๋‹ค. Hawk ๋…ผ๋ฌธ์€ ์‹ค์ œ ์‹œ์Šคํ…œ์˜ ์‘์šฉ๊ณผ ๋ฐฐํฌ๋ฅผ ๋‹ค๋ฃจ๋Š” ํ”„๋กœ๊ทธ๋žจ์— ํฌํ•จ๋๊ณ , Best Artifact Evaluation Award๋ฅผ ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค.

ํ•œ ๋ฌธ์žฅ์œผ๋กœ ๋ณด๋Š” Hawk: ADS131M08์ด ํŒŒํ˜•์„ ์ธก์ •ํ•˜๊ณ , ESP32๊ฐ€ ์ƒ˜ํ”Œ์„ ์ •๋ฆฌํ•˜๊ณ  ๋™๊ธฐํ™”ํ•˜๋ฉฐ, W5500์ด ์ˆ˜์ง‘ ๋ฐ์ดํ„ฐ๋ฅผ Ethernet์œผ๋กœ ์ „์†กํ•˜๊ณ , FFT-XGBoost๊ฐ€ ๊ฐ€์ „ ์ด๋ฒคํŠธ๋ฅผ ์ธ์‹ํ•ฉ๋‹ˆ๋‹ค.

Hawk ์‹คํ—˜์‹ค ๋ฐ์ดํ„ฐ ์ˆ˜์ง‘ ๋ณด๋“œ์™€ ์ „๊ธฐ ํŒจ๋„์— ์„ค์น˜๋œ ํ”„๋กœํ† ํƒ€์ž… ์ถœ์ฒ˜: Wang et al., Hawk, Figure 8, arXiv:2410.16293v1, CC BY 4.0.

NALM์ด๋ž€ ๋ฌด์—‡์ด๋ฉฐ ์ €์ „๋ ฅ ๊ฐ€์ „์€ ์™œ ์–ด๋ ค์šด๊ฐ€?

NALM์€ Non-Intrusive Appliance Load Monitoring, ์šฐ๋ฆฌ๋ง๋กœ ๋น„์นจ์ž…์‹ ๊ฐ€์ „ ๋ถ€ํ•˜ ๋ชจ๋‹ˆํ„ฐ๋ง์ž…๋‹ˆ๋‹ค. ์‹ค์ œ ์ธ์‹ ๋‹จ๊ณ„์—์„œ๋Š” ๊ฐ€์ „๋งˆ๋‹ค ์„ผ์„œ๋ฅผ ๊ณ„์† ์„ค์น˜ํ•˜์ง€ ์•Š๊ณ , ๋ถ„์ „๋ฐ˜์ด๋‚˜ ๋ฉ”์ธ ๋ฏธํ„ฐ ํ•œ ๊ณณ์—์„œ ํ•ฉ์ณ์ง„ ์ „๊ธฐ ๋ถ€ํ•˜๋ฅผ ์ธก์ •ํ•ด ์–ด๋–ค ๊ฐ€์ „์˜ ์ƒํƒœ๊ฐ€ ๋ฐ”๋€Œ์—ˆ๋Š”์ง€ ์ถ”์ •ํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๋Ÿฌ ์•…๊ธฐ๊ฐ€ ํ•จ๊ป˜ ์—ฐ์ฃผ๋˜๋Š” ์†Œ๋ฆฌ์—์„œ ์–ด๋–ค ์•…๊ธฐ๊ฐ€ ๋ง‰ ์—ฐ์ฃผ๋ฅผ ์‹œ์ž‘ํ•˜๊ฑฐ๋‚˜ ๋ฉˆ์ท„๋Š”์ง€ ๊ตฌ๋ถ„ํ•˜๋Š” ๊ฒƒ๊ณผ ๋น„์Šทํ•ฉ๋‹ˆ๋‹ค. Hawk๋Š” ํ•™์Šต์šฉ ๋ผ๋ฒจ์„ ๋งŒ๋“ค ๋•Œ๋งŒ ๊ฐœ๋ณ„ ๊ฐ€์ „ ์ „๋ฅ˜๋ฅผ ํ•จ๊ป˜ ์ธก์ •ํ•˜๊ณ , ํ˜„์žฅ ์ธ์‹์—์„œ๋Š” ๋ฉ”์ธ ๋ฏธํ„ฐ์˜ ํ•ฉ์‚ฐ ์ „๋ฅ˜ ์ŠคํŠธ๋ฆผ์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.

์ €์ „๋ ฅ ๊ธฐ๊ธฐ์˜ ๋ณ€ํ™”๋Š” ๋ƒ‰์žฅ๊ณ , ์—์–ด์ปจ ๊ฐ™์€ ํฐ ๋ฐฐ๊ฒฝ ๋ถ€ํ•˜์— ๋ฌปํžˆ๊ธฐ ์‰ฝ์Šต๋‹ˆ๋‹ค. Hawk๋Š” ํ•˜๋‚˜์˜ ํ‰๊ท  ์ „๋ ฅ๊ฐ’๋งŒ ๋ณด๋Š” ๋Œ€์‹  ์„ธ๋ฐ€ํ•œ ์ „๋ฅ˜ ํŒŒํ˜•์„ ์ˆ˜์ง‘ํ•˜๊ณ , ๊ฐ ๊ฐ€์ „์˜ ON/OFF ์ˆœ๊ฐ„์— ๋‚˜ํƒ€๋‚˜๋Š” ์ „๊ธฐ์  ํŒจํ„ด์„ ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค.

ํ”„๋กœ์ ํŠธ๋Š” ์„œ๋กœ ์ด์–ด์ง€๋Š” ๋‘ ๋‹จ๊ณ„๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. ๋จผ์ € ๊ท ํ˜• ์žˆ๊ฒŒ ๋ผ๋ฒจ๋ง๋œ ํŒŒํ˜• ๋ฐ์ดํ„ฐ์…‹์ธ HawkDATA๋ฅผ ๋งŒ๋“ค๊ณ , ๊ทธ๋‹ค์Œ ๋ฉ”์ธ ๋ฏธํ„ฐ์˜ ํ•ฉ์‚ฐ ์ „๋ฅ˜์—์„œ ๊ฐ€์ „ ์ด๋ฒคํŠธ๋ฅผ ์ธ์‹ํ•˜๋Š” ๋ชจ๋ธ์„ ํ•™์Šตํ•˜๊ณ  ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.

ESP32์™€ W5500 ๋ฐ์ดํ„ฐ ์ˆ˜์ง‘ ๋ณด๋“œ์˜ ๊ตฌ์กฐ

ADS131M08์€ ์ตœ๋Œ€ 8๊ฐœ ์•„๋‚ ๋กœ๊ทธ ์ฑ„๋„์„ 16 kHz, 24-bit ํ•ด์ƒ๋„๋กœ ๋™์‹œ์— ์ƒ˜ํ”Œ๋งํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. 50 Hz ์ „๋ ฅ ํ™˜๊ฒฝ์—์„œ ๊ต๋ฅ˜ ํ•œ ์ฃผ๊ธฐ๋Š” 20 ms์ด๋ฏ€๋กœ Hawk๋Š” ์ „์•• ํ•œ ์ฃผ๊ธฐ๋งˆ๋‹ค 320๊ฐœ์˜ ํŒŒํ˜• ์ ์„ ๊ธฐ๋กํ•ฉ๋‹ˆ๋‹ค. ์ด๋ ‡๊ฒŒ ์ด˜์ด˜ํ•˜๊ฒŒ ์ธก์ •ํ•˜๋ฉด ํ‰๊ท  ์ „๋ ฅ๊ฐ’๋งŒ์œผ๋กœ๋Š” ๋ณด๊ธฐ ์–ด๋ ค์šด ๊ณ ์กฐํŒŒ ํŒจํ„ด์„ ์‚ดํŽด๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

ESP32๋Š” ๋‘ ์ฝ”์–ด์— ์ˆ˜์ง‘ ์ž‘์—…์„ ๋‚˜๋ˆ•๋‹ˆ๋‹ค. Core1์€ ADC ์ธํ„ฐ๋ŸฝํŠธ์— ๋ฐ˜์‘ํ•ด ์ƒ˜ํ”Œ์„ ์ฝ๊ณ  ์ „์•• ์ฃผ๊ธฐ ๋‹จ์œ„๋กœ ๋ฌถ์Šต๋‹ˆ๋‹ค. Core0๋Š” ํŒจํ‚ท ์ƒํƒœ๋ฅผ ํ™•์ธํ•˜๊ณ  ์™„์„ฑ๋œ ๊ฐ ์ฃผ๊ธฐ์˜ ๋ฐ์ดํ„ฐ๋ฅผ ๋กœ๊น… ๊ฒฝ๋กœ๋กœ ๋ณด๋ƒ…๋‹ˆ๋‹ค. ๋ณด๋“œ์˜ ๋‘ ๋„คํŠธ์›Œํฌ ์ธํ„ฐํŽ˜์ด์Šค๋„ ์„œ๋กœ ๋‹ค๋ฅธ ์ผ์„ ๋งก์Šต๋‹ˆ๋‹ค.

  • ESP32 Wi-Fi๋Š” ๋ผ์šฐํ„ฐ์˜ TSF ์‹œ๊ฐ„ ์ •๋ณด๋ฅผ ์ฝ์Šต๋‹ˆ๋‹ค. Hawk๋Š” ๋ชจ๋“  ๋ณด๋“œ๊ฐ€ ํ•จ๊ป˜ ๊ด€์ธกํ•˜๋Š” ๊ต๋ฅ˜ ์ „์•• ์ฃผ๊ธฐ๋ฅผ shared perceptible time์œผ๋กœ ์‚ฌ์šฉํ•˜๊ณ , TSF, ์ฆ‰ Time Synchronization Function ๊ฐ’์œผ๋กœ ๊ฐ ์ฃผ๊ธฐ์˜ ID๋ฅผ ํ™•์ธํ•ฉ๋‹ˆ๋‹ค.
  • W5500 Ethernet์€ HawkDATA๋ฅผ ๋งŒ๋“œ๋Š” ๋™์•ˆ ํŒจํ‚ค์ง•๋œ ๊ณ„์ธก ๋ฐ์ดํ„ฐ๋ฅผ ๋ผ์šฐํ„ฐ๋ฅผ ๊ฑฐ์ณ ๋กœ๊น… PC๋กœ ์ „๋‹ฌํ•ฉ๋‹ˆ๋‹ค. ์ˆ˜์ง‘ ํ™˜๊ฒฝ๊ณผ ํ˜ธ์ŠคํŠธ ์‚ฌ์ด์—๋Š” MQTT๊ฐ€ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค.

Wi-Fi ํƒ€์ด๋ฐ๊ณผ W5500 Ethernet ์—ญํ• ์„ ๋ณด์—ฌ์ฃผ๋Š” Hawk ์ˆ˜์ง‘ ์•„ํ‚คํ…์ฒ˜ ์ถœ์ฒ˜: Wang et al., Hawk, Figure 7, arXiv:2410.16293v1, CC BY 4.0.

Figure 7์—์„œ ์ด ์—ญํ•  ๋ถ„๋ฆฌ๋ฅผ ๋ฐ”๋กœ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. Wi-Fi๋Š” ์‹œ๊ฐ„ ๋™๊ธฐํ™” ๊ฒฝ๋กœ์— ์ฐธ์—ฌํ•˜๊ณ , W5500์€ ๊ณ„์ธก ๋ฐ์ดํ„ฐ๋ฅผ ์˜ฎ๊ธฐ๋Š” ์œ ์„  ๊ฒฝ๋กœ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ESP32๊ฐ€ ์ƒ˜ํ”Œ๋ง๊ณผ ์ฃผ๊ธฐ ๋‹จ์œ„ ํŒจํ‚ค์ง•์„ ๋‹ด๋‹นํ•˜๋Š” ๋™์•ˆ ๋กœ๊ฑฐ๋Š” ๋ฐ์ดํ„ฐ์…‹์„ ๋งŒ๋“ค ํŒŒํ˜• ๊ธฐ๋ก์„ ์ „๋‹ฌ๋ฐ›์Šต๋‹ˆ๋‹ค.

๋ผ๋ฒจ์ด ๋ถ™์€ ํŒŒํ˜•์—์„œ HawkDATA๊นŒ์ง€

๋จธ์‹ ๋Ÿฌ๋‹ ํ•™์Šต์—๋Š” ํŒŒํ˜•๊ณผ ์ •๋‹ต ๊ฐ€์ „ ์ด๋ฒคํŠธ๊ฐ€ ์„œ๋กœ ๋งž๋Š” ์˜ˆ์ œ๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. Hawk๋Š” ๊ท ํ˜• Gray-code ์Šค์ผ€์ค„์„ ์ด์šฉํ•ด ๋ผ๋ฒจ์„ ๋งŒ๋“ญ๋‹ˆ๋‹ค. ํ•œ ๋‹จ๊ณ„๋งˆ๋‹ค ๋ชฉํ‘œ ๊ฐ€์ „ ํ•˜๋‚˜๋งŒ ์ƒํƒœ๋ฅผ ๋ฐ”๊พธ๊ธฐ ๋•Œ๋ฌธ์— '๊ฐ€์Šต๊ธฐ ON'์ด๋‚˜ '๋ชจ๋‹ˆํ„ฐ OFF'์ฒ˜๋Ÿผ ๋ช…ํ™•ํ•œ ์ด๋ฒคํŠธ๊ฐ€ ์ƒ๊น๋‹ˆ๋‹ค. ๊ฐ€์ „ ๊ทธ๋ฃน๊ณผ ์‹œ์ž‘ ์ƒํƒœ๋ฅผ ๋ฌด์ž‘์œ„๋กœ ๋ฐ”๊ฟ” ๋‹ค์–‘ํ•œ ๋ฐฐ๊ฒฝ ๋ถ€ํ•˜ ์กฐํ•ฉ๋„ ํ•™์Šต ๋ฐ์ดํ„ฐ์— ํฌํ•จํ•ฉ๋‹ˆ๋‹ค.

์‹คํ—˜์‹ค ๊ตฌ์„ฑ์—๋Š” ๋ชฉํ‘œ ๊ฐ€์ „ 18๊ฐœ์™€ ๋ฐฐ๊ฒฝ ๊ธฐ๊ธฐ 4๊ฐœ๋ฅผ ํ•ฉ์นœ 22๊ฐœ ๊ฐ€์ „์ด ์‚ฌ์šฉ๋์Šต๋‹ˆ๋‹ค. ์ „์ฒด raw ๋ฐ์ดํ„ฐ์…‹์€ 32.2์‹œ๊ฐ„, 5,796,650๊ฐœ ์ „์•• ์ฃผ๊ธฐ, 4,558๊ฐœ ๊ณ ์œ  ๊ฐ€์ „ ์ƒํƒœ ์กฐํ•ฉ, 127.8 GB์˜ ๋ฐ์ดํ„ฐ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค.

Hawk๋Š” ์—ฌ๋Ÿฌ ๋…ธ๋“œ์˜ ์ธก์ •๊ฐ’์„ ์„œ๋กœ ๋งž์ถ”๊ธฐ ์œ„ํ•ด shared perceptible time์ด๋ผ๋Š” ๋™๊ธฐํ™” ๋ฐฉ๋ฒ•๋„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋“  ๋…ธ๋“œ๊ฐ€ ๊ฐ™์€ ๊ต๋ฅ˜ ์ „์•• ์ฃผ๊ธฐ๋ฅผ ๊ด€์ธกํ•˜๊ณ , TSF ๊ฐ’์œผ๋กœ ๊ทธ ์ฃผ๊ธฐ๋ฅผ ์ผ๊ด€๋˜๊ฒŒ ์‹๋ณ„ํ•ฉ๋‹ˆ๋‹ค. ๋…ผ๋ฌธ์— ๋ณด๊ณ ๋œ ์ตœ๋Œ€ ๋™๊ธฐํ™” ์˜ค์ฐจ๋Š” 61.80 microseconds๋กœ, 16 kHz์—์„œ ๋‘ ์ƒ˜ํ”Œ ์‚ฌ์ด์˜ 62.5 microseconds๋ณด๋‹ค ์ž‘์Šต๋‹ˆ๋‹ค.

TSF์™€ ๋น„๊ตํ•œ Hawk ์ƒ˜ํ”Œ๋ง ๋™๊ธฐํ™” ์˜ค์ฐจ์˜ ๋ˆ„์ ๋ถ„ํฌ ์ถœ์ฒ˜: Wang et al., Hawk, Figure 10, arXiv:2410.16293v1, CC BY 4.0.

SsDiff, FFT, XGBoost๋Š” ์–ด๋–ป๊ฒŒ ์ด๋ฒคํŠธ๋ฅผ ์ธ์‹ํ•˜๋‚˜?

Hawk๊ฐ€ ์„ ํƒํ•œ ์ธ์‹ ๊ณผ์ •์€ ๋„ค ๋‹จ๊ณ„๋กœ ์ดํ•ดํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

  1. SsDiff๊ฐ€ ๋ณ€ํ™”๋ฅผ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค. SsDiff๋Š” Steady-state Differential current์˜ ์•ฝ์ž์ž…๋‹ˆ๋‹ค. ํ˜„์žฌ ํŒŒํ˜•์—์„œ 30๊ฐœ ๊ต๋ฅ˜ ์ฃผ๊ธฐ ์ „์˜ ํŒŒํ˜•์„ ๋บ๋‹ˆ๋‹ค. ๊ณ„์† ์ผœ์ ธ ์žˆ๋˜ ๋ฐฐ๊ฒฝ ๋ถ€ํ•˜๋Š” ๋Œ€๋ถ€๋ถ„ ์ƒ์‡„๋˜๊ณ , ๊ทธ์‚ฌ์ด ์ผœ์ง€๊ฑฐ๋‚˜ ๊บผ์ง„ ๊ฐ€์ „์˜ ๋ณ€ํ™”๊ฐ€ ๋” ๋ถ„๋ช…ํ•˜๊ฒŒ ๋‚จ์Šต๋‹ˆ๋‹ค.
  2. FFT๊ฐ€ ์ „๊ธฐ์  ์ง€๋ฌธ์„ ๋งŒ๋“ญ๋‹ˆ๋‹ค. FFT๋Š” Fast Fourier Transform, ๊ณ ์† ํ‘ธ๋ฆฌ์— ๋ณ€ํ™˜์ž…๋‹ˆ๋‹ค. ์‹œ๊ฐ„์— ๋”ฐ๋ผ ๋ณ€ํ•˜๋Š” ํŒŒํ˜•์„ ์–ด๋–ค ์ฃผํŒŒ์ˆ˜ ์„ฑ๋ถ„์ด ๋“ค์–ด ์žˆ๋Š”์ง€ ๋ณด์—ฌ์ฃผ๋Š” ํ˜•ํƒœ๋กœ ๋ฐ”๊ฟ‰๋‹ˆ๋‹ค. Hawk๋Š” ์ฒซ 10๊ฐœ ๊ณ ์กฐํŒŒ์˜ ์‹ค์ˆ˜๋ถ€, ํ—ˆ์ˆ˜๋ถ€, ํฌ๊ธฐ๋ฅผ ์‚ฌ์šฉํ•ด 320๊ฐœ ํŒŒํ˜• ์ ์„ 30๊ฐœ ํŠน์ง•์œผ๋กœ ์š”์•ฝํ•ฉ๋‹ˆ๋‹ค.
  3. XGBoost๊ฐ€ ์ง€๋ฌธ์„ ๋ถ„๋ฅ˜ํ•ฉ๋‹ˆ๋‹ค. XGBoost๋Š” ์•ž์„  ๊ฒฐ์ • ํŠธ๋ฆฌ์˜ ์˜ค๋ฅ˜๋ฅผ ๋’ค์˜ ํŠธ๋ฆฌ๊ฐ€ ๋ณด์™„ํ•˜๋„๋ก ์—ฌ๋Ÿฌ ํŠธ๋ฆฌ๋ฅผ ์ˆœ์ฐจ์ ์œผ๋กœ ๋งŒ๋“œ๋Š” gradient boosting ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. Hawk๋Š” ๋ชฉํ‘œ ๊ฐ€์ „ 18๊ฐœ์˜ ON/OFF 36๊ฐœ์™€ IDLE 1๊ฐœ๋ฅผ ํ•ฉ์นœ 37๊ฐœ ์ƒํƒœ ์ค‘ ํ•˜๋‚˜๋ฅผ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค.
  4. 30์ฃผ๊ธฐ ํˆฌํ‘œ๊ฐ€ ์ด๋ฒคํŠธ๋ฅผ ํ™•์ธํ•ฉ๋‹ˆ๋‹ค. ์งง์€ ๊ตฌ๊ฐ„์—์„œ ๊ฐ™์€ ์˜ˆ์ธก์ด ์ถฉ๋ถ„ํžˆ ๋ฐ˜๋ณต๋๋Š”์ง€ ํ™•์ธํ•ด ํ•œ ๋ฒˆ๋งŒ ๋‚˜ํƒ€๋‚˜๋Š” ์ž˜๋ชป๋œ ๊ฐ์ง€๋ฅผ ์ค„์ž…๋‹ˆ๋‹ค.

Hawk์˜ SsDiff, FFT-XGBoost ๋ถ„๋ฅ˜, ํˆฌํ‘œ ๊ธฐ๋ฐ˜ ์ถ”๋ก  ๊ณผ์ • ์ถœ์ฒ˜: Wang et al., Hawk, Figure 6, arXiv:2410.16293v1, CC BY 4.0.

ํ•™์Šต๊ณผ ์ถ”๋ก ์€ ์„œ๋กœ ๋‹ค๋ฅธ ๊ณณ์—์„œ ์ง„ํ–‰๋ฉ๋‹ˆ๋‹ค. ๋ชจ๋ธ์€ ๋ผ๋ฒจ์ด ๋ถ™์€ HawkDATA๋ฅผ ์‚ฌ์šฉํ•ด ์˜คํ”„๋ผ์ธ์—์„œ ํ•™์Šตํ•ฉ๋‹ˆ๋‹ค. ํ˜„์žฅ ์‹œ์Šคํ…œ์—์„œ๋Š” C๋กœ ์ž‘์„ฑ๋œ ์ถ”๋ก  ํ”„๋กœ๊ทธ๋žจ์ด Jetson Orin NX 16GB์—์„œ ์‹คํ–‰๋˜๊ณ , MQTT๋กœ ๋ฉ”์ธ ๋ฏธํ„ฐ์˜ ํ•ฉ์‚ฐ ์ „๋ฅ˜ ์ŠคํŠธ๋ฆผ์„ ๊ตฌ๋…ํ•ฉ๋‹ˆ๋‹ค. ๋…ผ๋ฌธ์€ ์ด ํ™˜๊ฒฝ์—์„œ ํ‰๊ท  ์ถ”๋ก  ์ง€์—ฐ 4.07 ms์™€ ๋ฉ”๋ชจ๋ฆฌ ์‚ฌ์šฉ๋Ÿ‰ 40.19 MB๋ฅผ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. ESP32์™€ W5500์€ ์ˆ˜์ง‘๊ณผ ํ†ต์‹ ์„ ๋‹ด๋‹นํ•˜๋Š” ์•ž๋‹จ์ด๊ณ , Jetson์ด ํ•™์Šต๋œ ์ธ์‹ ๋ชจ๋ธ์„ ์‹คํ–‰ํ•ฉ๋‹ˆ๋‹ค.

๋ฐ์ดํ„ฐ์…‹๊ณผ ์‹ค์ œ ๊ณต๊ฐ„์—์„œ ์–ป์€ ๊ฒฐ๊ณผ

Hawk๋Š” F1 ์ ์ˆ˜๋กœ ๊ฒฐ๊ณผ๋ฅผ ๋ณด๊ณ ํ•ฉ๋‹ˆ๋‹ค. F1์€ ์ •๋ฐ€๋„์™€ ์žฌํ˜„์œจ์„ ํ•จ๊ป˜ ๋ณด๋Š” ์ง€ํ‘œ์ž…๋‹ˆ๋‹ค. ์‰ฝ๊ฒŒ ๋งํ•ด ์‹ค์ œ ๊ฐ€์ „ ์ด๋ฒคํŠธ๋ฅผ ์ž˜ ์ฐพ์œผ๋ฉด์„œ ์ž˜๋ชป๋œ ์ด๋ฒคํŠธ ๋ณด๊ณ ๋Š” ์ค„์—ฌ์•ผ ๋†’์•„์ง‘๋‹ˆ๋‹ค. ๋…ผ๋ฌธ์€ ํ‰๊ฐ€ ๋Œ€์ƒ๋ณ„ ๊ฒฐ๊ณผ๋ฅผ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค.

  • HawkDATA ์ด๋ฒคํŠธ ์ธ์‹ ํ‰๊ท  F1 92.71%
  • HawkDATA ์ƒํƒœ ์ธ์‹ F1 ์•ฝ 93.94%
  • ๋ณ„๋„ BLUED ๋ฐ์ดํ„ฐ์…‹ ์ด๋ฒคํŠธ ๋ถ„๋ฅ˜ F1 97.07%
  • ์ฃผ๊ฑฐ ๊ณต๊ฐ„๊ณผ ์‚ฌ๋ฌด์‹ค ํ˜„์žฅ ํ‰๊ฐ€ ์ด๋ฒคํŠธ F1 ๊ฐ๊ฐ 96.02%, 94.76%

ํ˜„์žฅ ํ‰๊ฐ€๋Š” 103 mยฒ ์ฃผ๊ฑฐ ๊ณต๊ฐ„๊ณผ 25 mยฒ ๋Œ€ํ•™์›์ƒ ์‚ฌ๋ฌด์‹ค์—์„œ ์ง„ํ–‰๋์œผ๋ฉฐ, ์•Œ ์ˆ˜ ์—†๋Š” ๋ฐฐ๊ฒฝ ๊ฐ€์ „๋„ ํ•จ๊ป˜ ๋™์ž‘ํ–ˆ์Šต๋‹ˆ๋‹ค. ์•„๋ž˜ ํ‰๋ฉด๋„๋Š” ๋ฉ”์ธ ๋ฏธํ„ฐ ์„ผ์„œ์™€ ํ‰๊ฐ€ ๋Œ€์ƒ ๊ฐ€์ „์ด ์–ด๋””์— ๋ฐฐ์น˜๋๋Š”์ง€ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.

Hawk ์ฃผ๊ฑฐ ๊ณต๊ฐ„๊ณผ ์‚ฌ๋ฌด์‹ค ๋ฐฐํฌ ํ‰๋ฉด๋„ ์ถœ์ฒ˜: Wang et al., Hawk, Figure 9, arXiv:2410.16293v1, CC BY 4.0.

โš™๏ธ W5500์˜ ์—ญํ• ์ด ์ค‘์š”ํ•œ ์ด์œ 

Hawk๋Š” W5500์ด ๋จธ์‹ ๋Ÿฌ๋‹ ์—ฐ์‚ฐ์„ ์ง์ ‘ ์ˆ˜ํ–‰ํ•˜์ง€ ์•Š๋”๋ผ๋„ ML ์‹œ์Šคํ…œ์„ ์–ด๋–ป๊ฒŒ ๋’ท๋ฐ›์นจํ•  ์ˆ˜ ์žˆ๋Š”์ง€ ๋ณด์—ฌ์ฃผ๋Š” ์‚ฌ๋ก€์ž…๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ์…‹ ๊ตฌ์ถ• ๋‹จ๊ณ„์—์„œ W5500์€ ์ „์•• ์ฃผ๊ธฐ ๋‹จ์œ„๋กœ ํŒจํ‚ค์ง•๋œ ๊ณ ํ•ด์ƒ๋„ ๊ณ„์ธก ๋ฐ์ดํ„ฐ๋ฅผ ESP32 ์ˆ˜์ง‘ ๋ณด๋“œ์—์„œ ๋กœ๊น… PC๊นŒ์ง€ ์ „๋‹ฌํ•ฉ๋‹ˆ๋‹ค. ์ด๋ ‡๊ฒŒ ๋ชจ์ธ ๊ธฐ๋ก์ด ๊ฐ€์ „ ์ธ์‹ ๋ชจ๋ธ์„ ๊ฐœ๋ฐœํ•˜๊ณ  ํ‰๊ฐ€ํ•˜๋Š” ๋ผ๋ฒจ ๋ฐ์ดํ„ฐ๊ฐ€ ๋ฉ๋‹ˆ๋‹ค.

๊ตฌ์„ฑ ์š”์†Œ๋ณ„ ์—ญํ• ์€ ํ•˜๋‚˜์˜ ์—ฐ๊ฒฐ๋œ ํ๋ฆ„์„ ๋งŒ๋“ญ๋‹ˆ๋‹ค.

  • ADS131M08์€ ์•„๋‚ ๋กœ๊ทธ ์ „๋ฅ˜์™€ ์ „์•• ์‹ ํ˜ธ๋ฅผ ๋™๊ธฐํ™”๋œ ๋””์ง€ํ„ธ ์ƒ˜ํ”Œ๋กœ ๋ฐ”๊ฟ‰๋‹ˆ๋‹ค.
  • ESP32๋Š” ์ˆ˜์ง‘, ์ „์•• ์ฃผ๊ธฐ ๋‹จ์œ„ ํŒจํ‚ค์ง•, ์‹œ๊ฐ„ ์ •๋ณด ์กฐํšŒ๋ฅผ ๋‹ด๋‹นํ•ฉ๋‹ˆ๋‹ค.
  • W5500์€ ๊ณ„์ธก ๋ฐ์ดํ„ฐ์˜ ์œ ์„  Ethernet ๊ฒฝ๋กœ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
  • ๋กœ๊น… ํ™˜๊ฒฝ์€ HawkDATA๋ฅผ ์ €์žฅํ•˜๊ณ  ์ •๋ฆฌํ•ฉ๋‹ˆ๋‹ค.
  • Jetson Orin NX๋Š” ๋ฉ”์ธ ๋ฏธํ„ฐ ์ŠคํŠธ๋ฆผ์—์„œ ์‹ค์‹œ๊ฐ„ ์ถ”๋ก ์„ ์‹คํ–‰ํ•ฉ๋‹ˆ๋‹ค.

WIZnet ๊ฐœ๋ฐœ์ž ๊ด€์ ์—์„œ ์žฌ์‚ฌ์šฉํ•  ๋งŒํ•œ ์„ค๊ณ„ ํฌ์ธํŠธ๋Š” ์ž„๋ฒ ๋””๋“œ ์„ผ์‹ฑ๊ณผ ๋จธ์‹ ๋Ÿฌ๋‹ ์‚ฌ์ด์˜ ์—ฐ๊ฒฐ์ž…๋‹ˆ๋‹ค. ๋„คํŠธ์›Œํฌ ์ปจํŠธ๋กค๋Ÿฌ๊ฐ€ ๊ณ ์† ๋ฐ์ดํ„ฐ ์ˆ˜์ง‘ ์žฅ์น˜์™€ ๋ฐ์ดํ„ฐ์…‹์„ ๋งŒ๋“ค๊ณ  ๋ชจ๋ธ์„ ์‹คํ–‰ํ•˜๋Š” ํ˜ธ์ŠคํŠธ๋ฅผ ์ด์–ด์ค„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

๊ณต๊ฐœ๋œ ์—ฐ๊ตฌ ์ž๋ฃŒ

์—ฐ๊ตฌํŒ€์€ Hawk ๋…ผ๋ฌธ, SenSys24-Hawk artifact ์ €์žฅ์†Œ, HawkDATA ํ‰๊ฐ€ ํŒจํ‚ค์ง€๋ฅผ ๊ณต๊ฐœํ–ˆ์Šต๋‹ˆ๋‹ค. ์ €์žฅ์†Œ์—๋Š” ํ‰๊ฐ€ ์Šคํฌ๋ฆฝํŠธ, XGBoost ๋ชจ๋ธ, BLUED ์ด๋ฒคํŠธ ๋ฐ์ดํ„ฐ, ๋™๊ธฐํ™” ๋ถ„์„, ์งง์€ ํ˜„์žฅ ์˜์ƒ์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. SenSys 2024๋Š” ์ด ์—ฐ๊ตฌ ์ž๋ฃŒ์— Best Artifact Evaluation Award๋ฅผ ์ˆ˜์—ฌํ–ˆ์Šต๋‹ˆ๋‹ค.

์ด ์ž๋ฃŒ๋กœ ๋ฐ์ดํ„ฐ์…‹, ๋™๊ธฐํ™” ๋ฐฉ์‹, ๋ชจ๋ธ ํ‰๊ฐ€, ํ˜„์žฅ ๊ฒฐ๊ณผ๋ฅผ ์‚ดํŽด๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ž„๋ฒ ๋””๋“œ ์ˆ˜์ง‘ ๋ณด๋“œ๋ฅผ ์‹ค์ œ๋กœ ์ œ์ž‘ํ•˜๋ ค๋ฉด artifact release์— ํฌํ•จ๋˜์ง€ ์•Š์€ ์ˆ˜์ง‘ ํŽŒ์›จ์–ด, ํšŒ๋กœ๋„, PCB, BOM, broker ์„ค์ •์„ ๋ณ„๋„๋กœ ์„ค๊ณ„ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.

๊ด€๋ จ WIZnet Maker ์ฝ์„๊ฑฐ๋ฆฌ

์—ฌ๋Ÿฌ ์ฑ„๋„์˜ ๋ฐ์ดํ„ฐ๋ฅผ AI ์›Œํฌํ”Œ๋กœ๋กœ ๋„˜๊ธฐ๋Š” ๋˜ ๋‹ค๋ฅธ ์‚ฌ๋ก€๋กœ How Does W5500 Stream 8-Channel RF Data for Drone Detection?๋ฅผ ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ํ”„๋กœ์ ํŠธ๋Š” RF ๊ฐ์ง€๋ฅผ ์œ„ํ•ด STM32์™€ AD7606์„ ์‚ฌ์šฉํ•˜๊ณ , Hawk๋Š” ์ „๋ ฅ ํŒŒํ˜•์„ ์œ„ํ•ด ESP32์™€ ADS131M08์„ ์‚ฌ์šฉํ•œ๋‹ค๋Š” ์ฐจ์ด๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค.

TheBreeze: W5500 Ethernet DDP Receiver for WLED๋Š” LED ์‹œ์Šคํ…œ์—์„œ ๋น„์Šทํ•œ ์œ ๋ฌด์„  ์—ญํ•  ๋ถ„๋ฆฌ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. W5500์€ ์ฃผ๋œ pixel stream์„ ์ฒ˜๋ฆฌํ•˜๊ณ  Wi-Fi๋Š” ๋” ๊ฐ€๋ฒผ์šด ์ œ์–ด์— ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค.

๋ณด๋‹ค ์‹ค์šฉ์ ์ธ ์—๋„ˆ์ง€ gateway ์‚ฌ๋ก€๋Š” ESP32 JSY Energy Meter Gateway with W5500 Support์ž…๋‹ˆ๋‹ค. ์ด ์žฅ์น˜๋Š” Hawk์˜ ๊ณ ํ•ด์ƒ๋„ ํ•™์Šต ํŒŒํ˜• ๋Œ€์‹  ๊ฐ€๊ณต๋œ ์ „๋ ฅ๊ณ„ telemetry๋ฅผ ์ „์†กํ•ฉ๋‹ˆ๋‹ค. ๋‘ ํ”„๋กœ์ ํŠธ๋ฅผ ํ•จ๊ป˜ ๋ณด๋ฉด ์„ค์น˜ํ˜• ์—๋„ˆ์ง€ gateway์™€ ํ•™์ˆ  ML ๋ฐ์ดํ„ฐ ํŒŒ์ดํ”„๋ผ์ธ์˜ ์ฐจ์ด๋ฅผ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

์ถœ์ฒ˜ ๊ธฐ๋ฐ˜ ์š”์•ฝ

Hawk๋Š” ์ž„๋ฒ ๋””๋“œ ๊ณ„์ธก๊ณผ ๊ฐ€์ „ ์ธ์‹ ๋จธ์‹ ๋Ÿฌ๋‹์„ ๋ถ„๋ช…ํ•œ ์ˆœ์„œ๋กœ ์—ฐ๊ฒฐํ•ฉ๋‹ˆ๋‹ค. ADS131M08๋กœ ์ˆ˜์ง‘ํ•˜๊ณ , ESP32๋กœ ์ •๋ฆฌํ•˜๊ณ  ๋™๊ธฐํ™”ํ•˜๋ฉฐ, W5500์œผ๋กœ ์ „์†กํ•˜๊ณ , FFT๋กœ ๊ณ ์กฐํŒŒ ํŠน์ง•์„ ์ถ”์ถœํ•˜๊ณ , XGBoost๋กœ ๋ถ„๋ฅ˜ํ•œ ๋’ค, 30์ฃผ๊ธฐ ํˆฌํ‘œ๋กœ ์ด๋ฒคํŠธ๋ฅผ ํ™•์ •ํ•ฉ๋‹ˆ๋‹ค. ๋…ผ๋ฌธ, ๋ฐ์ดํ„ฐ์…‹, Best Artifact Evaluation Award๋ฅผ ๋ฐ›์€ ์—ฐ๊ตฌ ์ž๋ฃŒ, ํ˜„์žฅ ํ‰๊ฐ€๊ฐ€ ํ•จ๊ป˜ ๊ณต๊ฐœ๋œ Ethernet ๊ธฐ๋ฐ˜ AIoT ๋ฐ์ดํ„ฐ ์ˆ˜์ง‘ ์—ฐ๊ตฌ ์‚ฌ๋ก€์ž…๋‹ˆ๋‹ค.

โ“ FAQ

Q. NALM์€ ๋ฌด์Šจ ๋œป์ธ๊ฐ€์š”? NALM์€ Non-Intrusive Appliance Load Monitoring, ๋น„์นจ์ž…์‹ ๊ฐ€์ „ ๋ถ€ํ•˜ ๋ชจ๋‹ˆํ„ฐ๋ง์ž…๋‹ˆ๋‹ค. ๊ฐ€์ „๋งˆ๋‹ค ์„ผ์„œ๋ฅผ ๋‹ฌ์ง€ ์•Š๊ณ  ๋ฉ”์ธ ๋ฏธํ„ฐ์˜ ํ•ฉ์‚ฐ ์ธก์ •๊ฐ’์œผ๋กœ ๊ฐœ๋ณ„ ๊ฐ€์ „์˜ ๋™์ž‘์„ ์ถ”์ •ํ•ฉ๋‹ˆ๋‹ค.

Q. FFT์™€ XGBoost๋Š” ๊ฐ๊ฐ ๋ฌด์—‡์„ ํ•˜๋‚˜์š”? FFT๋Š” ์ „๋ฅ˜ ํŒŒํ˜•์„ ๊ณ ์กฐํŒŒ ํŠน์ง•์œผ๋กœ ๋ฐ”๊พธ๊ณ , XGBoost๋Š” ๊ทธ ํŠน์ง•์„ ๋ณด๊ณ  ์–ด๋–ค ๊ฐ€์ „์ด ์ผœ์ง€๊ฑฐ๋‚˜ ๊บผ์กŒ๋Š”์ง€ ๋ถ„๋ฅ˜ํ•ฉ๋‹ˆ๋‹ค.

Q. Wi-Fi์™€ W5500์€ ๊ฐ๊ฐ ์–ด๋–ค ์—ญํ• ์ธ๊ฐ€์š”? ESP32 Wi-Fi๋Š” ๋™๊ธฐํ™”๋ฅผ ์œ„ํ•œ TSF ์‹œ๊ฐ„ ์ •๋ณด๋ฅผ ์ฝ์Šต๋‹ˆ๋‹ค. W5500์€ ๋ฐ์ดํ„ฐ์…‹ ๊ตฌ์ถ• ๋‹จ๊ณ„์—์„œ ํŒจํ‚ค์ง•๋œ ์ˆ˜์ง‘ ๋ฐ์ดํ„ฐ๋ฅผ ์œ ์„  Ethernet์œผ๋กœ ๋กœ๊น… PC์— ์ „๋‹ฌํ•ฉ๋‹ˆ๋‹ค.

Q. ํ•™์Šต๊ณผ ์‹ค์‹œ๊ฐ„ ์ถ”๋ก ์€ ์–ด๋””์—์„œ ์ง„ํ–‰๋˜๋‚˜์š”? ํ•™์Šต์€ ๋ผ๋ฒจ์ด ๋ถ™์€ HawkDATA๋กœ ์˜คํ”„๋ผ์ธ์—์„œ ์ง„ํ–‰๋ฉ๋‹ˆ๋‹ค. ๋ฐฐํฌ๋œ C ์ถ”๋ก  ํ”„๋กœ๊ทธ๋žจ์€ Jetson Orin NX 16GB์—์„œ ์‹คํ–‰๋˜๋ฉฐ MQTT๋กœ ๋ฉ”์ธ ๋ฏธํ„ฐ ์ŠคํŠธ๋ฆผ์„ ์ฝ์Šต๋‹ˆ๋‹ค.

Q. artifact ์ €์žฅ์†Œ๋งŒ์œผ๋กœ ์ˆ˜์ง‘ ๋ณด๋“œ ์ „์ฒด๋ฅผ ์ œ์ž‘ํ•  ์ˆ˜ ์žˆ๋‚˜์š”? ์ €์žฅ์†Œ๋ฅผ ํ†ตํ•ด ์•Œ๊ณ ๋ฆฌ์ฆ˜๊ณผ ํ‰๊ฐ€ ๊ณผ์ •์„ ์‚ดํŽด๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ „์ฒด ํ•˜๋“œ์›จ์–ด ์ œ์ž‘์—๋Š” release์— ํฌํ•จ๋˜์ง€ ์•Š์€ ์ˆ˜์ง‘ ํŽŒ์›จ์–ด์™€ ๋ณด๋“œ ์„ค๊ณ„ ์ž๋ฃŒ๊ฐ€ ์ถ”๊ฐ€๋กœ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค.

Documents

  • Hawk paper - arXiv HTML

    Primary CC BY 4.0 paper source with Figures 7-10 and W5500 acquisition-board evidence

  • Hawk paper - arXiv PDF

    Downloadable paper version

  • Official SenSys 2024 program

    Hawk in the Applications and Deployments paper session

  • Official SenSys 2024 awards

    Hawk received the Best Artifact Evaluation Award

  • Hawk artifact repository

    Evaluation scripts, XGBoost models, BLUED data, synchronization analysis, and field video; repository has no LICENSE

  • HawkDATA dataset

    Public evaluation dataset package referenced by the artifact repository

  • WIZnet W5500

    Official product page for the Ethernet controller used in the Hawk acquisition board

Comments

Similar projects you might like

Comments