Machine learning classification of frequency-hopping spread spectrum signals in a multi-signal environment

Abstract

Frequency-hopping spread spectrum (FHSS) spreads the signal over a wide bandwidth, where the carrier frequencies change rapidly according to a pseudorandom number making signal classification difficult. Classification becomes more complex with the presence of additive white Gaussian noise (AWGN) and interference due to background signals. In this research, a hybrid convolutional neural network (HCNN) system with the fusion of handcrafted and deep features is proposed to classify FHSS signals in the presence of AWGN and the background signal. The CNN is used as a deep feature extractor by transforming the intermediate frequency (IF) signal to the time-frequency representation (TFR) and used as a two-dimensional (2D) input image, whereas the handcrafted features of the FHSS signal such as hop frequency and hop duration are estimated from the TFR. A proper network structure of the three-layer fully connected network (TLFCN) is determined and used as a classifier. The TLFCN is a machine learning algorithm that requires training with a proper dataset to classify the various types of FHSS signals. Ideally, the dataset size must be sufficiently large as well as balanced to optimize the classification performance. A pseudorandom sequence of hopping frequencies observed from an FHSS signal represents one observation of all the possible hopping sequences of the signal. Therefore, an observation calculating technique is developed that can derive the total number of possible hopping sequences of an FHSS signal by using the frequencies to determine the observations in the dataset. The majority of the machine learning algorithms assume that the training set is evenly distributed among classes. However, in many real-world applications, the number of observations among classes is often imbalanced, which reduces the classification performance of the algorithm. The number of observations of an FHSS signal depends on the number of hop frequencies. Therefore, a given set of FHSS signals with a varying number of hop frequencies among the FHSS signals results in an uneven number of observations, thereby building an imbalanced dataset. Thus, resampling and data augmentation methods such as synthetic minority oversampling technique (SMOTE) and random erasing (RE) are performed to balance the dataset for the increased learning and decision-making capacity of a machine learning algorithm. Monte Carlo simulation is performed to verify the classification performance of the linear discriminant (LD), TLFCN, CNN, and HCNN for various signal-to-noise ratio (SNR) levels. Based on the SNR range at 90% probability of correct classification (PCC) in the presence of AWGN and the background signal, the LD performed worst from 1 to 15 dB among all the methods, whereas the HCNN performed best from -1.58 to -0.66 dB. Moreover, the HCNN with the balanced dataset performed better by 0.14 to 1.06 dB of SNR than with the imbalanced dataset. Therefore, the HCNN system improved the classification performance and performed better than conventional machine learning-based algorithms.