TK1109 : Modulation Classification Using Machine Learning Methods
Thesis > Central Library of Shahrood University > Electrical Engineering > MSc > 2026
Authors:
Abstarct: Abstract
Automatic Modulation Classification (AMC) identifies the modulation format of a received waveform when transmitter-side signaling information is unavailable. This thesis develops and evaluates an interpretable, feature-baxsed machine learning frxamework for five digital modulation classes: BPSK, QPSK, 8PSK, 16-QAM, and 64-QAM. A controlled experiment in MATLAB comprises 7,000 independently generated frxames across seven SNR levels ranging from −10 to 20 dB and under two channel conditions: AWGN and flat block Rayleigh fading.
Each frxame is centered and power-normalized prior to feature extraction. The baxseline representation consists of a 22-dimensional feature vector comprising amplitude statistics, moments and complex cumulants, phase descxriptors, rotation-invariant circular harmonics, and spectral measures. baxseline models use a common stratified 70/30 train-test split. The empirical study was then strengthened using ten repeated test splits, feature-family ablation analysis, and five-fold cross-validation performed exclusively on the training data for all 31 non-empty combinations of the five feature families. The One-SE rule selected a compact 10-feature representation consisting of seven amplitude descxriptors and three complex statistical features.
In the repeated-split stability analysis using the 22-feature baxseline, Random Forest achieved a mean accuracy of 65.26%, SVM achieved 64.87%, and KNN achieved 59.89%; the RF-SVM difference was not statistically distinguishable at the 0.05 significance level (p = 0.124). Training-only feature optimization selected the 10 features without using labels from the held-out test set. In the final fixed-test comparison, SVM with the selected 10 features achieved an accuracy of 65.19% and a Macro-F1 score of 66.09%, while Random Forest achieved 64.67% and 65.06%, and KNN achieved 63.62% and 63.95%, respectively. The 10-feature representation reduced SVM training and inference time while maintaining or improving classification performance. Performance remained strongly dependent on SNR, with confusion between 16-QAM and 64-QAM constituting the most significant remaining source of error. Therefore, the final engineering recommendation is to use SVM with the 10-feature representation baxsed on amplitude and complex statistical features.
Keywords:
#_Automatic Modulation Classification #Machine Learning #SVM #Random Forest #KNN #Digital Modulation #AWGN #Rayleigh Fading #MATLAB_ Keeping place: Central Library of Shahrood University
Visitor:
Visitor: