| dc.description.abstract | This research aims to identify Amplitude Shift Keying (ASK) modulation signals
using the MobileNetV2 architecture with a modified activation function approach.
MobileNetV2 was selected due to its high computational efficiency and suitability for
resource-constrained systems. However, this architecture has a limitation in terms of
classification accuracy. To address this, experiments were conducted by replacing the
default ReLU6 activation function with alternative functions Swish and Mish to improve
model performance. The model was trained using a dataset from the Kaggle platform,
which consists of ASK, BPSK, and QPSK signals across various Signal-to-Noise Ratio
(SNR) levels. Training was conducted for 10, 20, and 30 epochs for each activation
function. The results of the 30-epoch training showed that the Mish activation function
achieved the highest validation accuracy of 91.53%, followed by Swish at 91.29%, and
ReLU6 at 90.48%. Mish also demonstrated stable learning behavior despite requiring the
longest training time. Swish yielded the lowest validation loss of 0.1755, indicating
efficient and stable model performance. Meanwhile, ReLU6 experienced training
anomalies due to its limited activation range. Overall, replacing the activation function has
been proven to significantly improve ASK signal classification performance in the
MobileNetV2 architecture. | en_US |
