51. Events identification in the Internet of Things environment by applying machine learning
- Author
-
Leto, Ivana and Jurčević, Marko
- Subjects
ugradbeni uređaji ,TinyML ,machine learning ,internet stvari ,TECHNICAL SCIENCES. Computing ,embedded devices ,TEHNIČKE ZNANOSTI. Računarstvo ,Internet of Things ,prepoznavanje pokreta ,movement recognition ,strojno učenje - Abstract
TinyML je novi koncept strojnog učenja koji omogućuje prikupljanje podataka za učenje, treniranje i testiranje modela na malim uređajima snage u rasponu od samo nekoliko milivata. Takav koncept omogućava veliku raznolikost i laku primjenjivost. Prednosti tehnologija koji ga omogućavaju su mali troškovi, mala potrošnja energije, mala latencija, velika sigurnost podataka i neovisnost o povezanosti na internet.TinyML je isto tako i suočen s mnogim izazovima zbog čega nije idealno rješenje za sve sustave. Samo neki od problema su upravo i neke od prednosti, naime zbog velike energetske učinkovitosti koju imaju uređaji pogodni za TinyML snaga procesiranja podataka za strojno učenje je ograničena čime je najčešće i točnost modela narušena. Osim pojašnjenja samog TinyML-a i navođenja nekih od sklopovlja i razvojnih okvira koji ga podržavaju, u ovom radu je obrađena i praktična primjena TinyML-a na pločici FiPy uz Pysense pločicu sa senzorima za detekciju pokreta uređaja. Za programiranje je korišten MicroPython pomoću kojeg su prikupljeni podatci za učenje te klasificiranje novih uzoraka korištenjem KNN algoritma. Povezivanjem uređaja na Wi-Fi novi uzorci su poslani i spremljeni u MySQL bazu s njihovim predviđenim klasama te grafički prikazani alatom Grafana. TinyML is a new machine learning concept that allows data collection for learning, training and model testing on small power devices ranging from just a few milliwatts. Such a concept allows for great variety and easy applicability. Advantages of such technology are low costs, low power consumption, low latency, high data security and independence of internet connection. TinyML also faces many challenges which is why it is not an ideal solution for all systems. Some of the problems are also some of the advantages of the concept, namely due to the high energy efficiency of devices suitable for TinyML, the power of data processing for machine learning is limited, which most often impairs the accuracy of the model. In addition to explaining TinyML itself and listing some of the circuits and development frameworks that support it, this paper also discusses the practical application of TinyML on a FiPy board alongside a Pysense board with sensors to detect device movements. MicroPython was used for programming the device. It was used to collect learning data and classify new samples using KNN algorithm. Connecting the device to Wi-Fi, the new samples are sent and stored to the MySQL database together with its predicted classes and graphically displayed with the Grafana tool.
- Published
- 2022