Resumo:
This work investigates an artificial intelligence-based method applied to edge processing
for the recognition of water meter digits. The implementation includes two low-power
transmitters, NBIoT and LoRa, for transmitting the recognized numbers, and estimates
the device’s energy consumption to evaluate the feasibility of low-power solutions in edge
processing applications.
A total of 387 image recognition and edge processing measurements were conducted with
transmission via NBIoT, and 241 measurements with LoRa. The edge processing results,
obtained from the training of a neural network developed in this work, showed an average
error of 0.289 and an accuracy of 93.52%, indicating high reliability in identifying the
water meter’s numerical values.
Data transmission via NBIoT showed low packet loss (4.65%), while transmission with
LoRa showed no loss, demonstrating high precision in data transmission. The application
of edge processing with battery and low-power transmitters proved promising for use in
water meters with daily measurements, estimating an autonomy of 623 days for LoRa
applications and 631 days for NBIoT. Thus, the feasibility of low-power solutions in edge
processing applications is confirmed.
