Page 44 - Kỷ yếu hội thảo khoa học lần thứ 12 - Công nghệ thông tin và Ứng dụng trong các lĩnh vực (CITA 2023)

28

We imported the Excel file and used pandas Python Data Analysis Library to read
and save into correct variable. Then, we converted these sounds to spectrogram with a
size of 30x32 as input. The model will be trained eventually with the support of Keras

Embedding model into STM32: In this research, we used STM32F746NGH6 Discov-
ery kit for main microcontroller to run this model (Fig. 6). This Discovery Kit is suitable
to run Edge AI and has a rich number of peripherals that allow communication to read
many types of sensors. Moreover, STMicroelectronics proposes X-CUBE-AI - a soft-
ware package that makes implementation of deep-learning models on microcontrollers
easier [12].
For input collection, SparkFun MEMS-Microphone AMDP401 can be used
low-cost, analog output microphone that is small and has a low current consumption
only 250uA and a wide-band frequency ranged from 100 Hz to15 kHz, making it a
good choice for embedded applications.
converted by using ADC channel 3 of the STM32 with sampling frequency 16 kHz.

crocontroller using vendor-product Cube-MX AI.

Fig. 6. Microcontroller STM32 Kit and connection with MEMS-microphone

3.2 Results

The total training time was 26 minutes and 40 seconds with 423 samples. The model
loss is show in Fig. 7. The model loss, representing the difference between the model
result (prediction) and the desired outcome (goal), approaches to 0 in later epoch,
proved acceptable performance of the proposed model.

CITA 2023 ISBN: 978-604-80-8083-9