This is an assignment for Pattern Recognition Course taught at Alexandria University, Faculty of Engineering offered in Spring 2019. The assignment goal is to design neural network that are able to classify the signals in the DeepSig dataset into their different modulation types.
DeepSig Dataset: RadioML 2016.04C
A synthetic dataset, generated with GNU Radio, consisting of 11 modulations. This is a
variable-SNR dataset with moderate LO drift, light fading, and numerous different
labeled SNR increments for use in measuring performance across different signal and
noise power scenarios.
Data version used is available from here.
Every sample is presented using two vectors each of them has 128 elements. The dataset has the signals with raw features so we extracted more features
from it which are:
○ First Derivative in Time
○ Integral in Time
We used extracted features and created a combination of them in order to
have more datasets with different features to train the model, the
combinations are:
○ Derivative Features + Raw Features
○ Integral Features + Raw Features
○ Integral Features + Derivative Features
○ Integral Features + Derivative Features + Raw Features
As the size of datasets with the combination of features was very large to fit b in the memory, we converted them to T
ensorflow Records so we can train the model with data loaded on disk. Each dataset is splitted into 50% for training/validation and 50% for testing
Architecture used for the FCN is:
○ Input layer of size (2,128)
○ First hidden layer of size 512
○ Second hidden layer of size 512
○ Flatten Layer
○ Softmax output layer of size 10
Activation function of neurons is ReLU. The optimizer used is Adam. The loss function used is Categorical Cross Entropy.
Results:
Best model from this approach was the one using the first derivative in time and raw features combination and scored a 45.86% testing accuracy in the evaluation.
Architecture used for CNN:
○ Input layer of size (128,2,1)
○ Convolutional layer with 64 filters of size (3,1)
○ Max Pooling Layer
○ Convolutional Layer with 16 filters of size (3,2)
○ Flatten Layer
○ Dropout Layer
○ Fully connected layer of size 128
○ Softmax output Layer
Results:
Best model fromn this approach was the one using the integral in time features and scored 56.46% testing accuracy in evaluation.
More results and diagrams comprising the accuracy and realtion with respect to SNR can be found in the Evalution notebook
1- T. O’shea, N. West “Radio Machine Learning Dataset Generation with
GNU Radio”
2-T. O’Shea, J. Corgan, and T. Clancy “Convolutional Radio
Modulation Recognition Networks
3-N. West, T. O’shea “Deep Architectures for Modulation Recognition”
5-K. Karra, S. Kuzdeba, J. Peterson “Modulation recognition using
hierarchical deep neural networks”