Automatic Modulation Classification Using Interacting Multiple Model Kalman Filter for Channel Estimation

Abstract

T 0018-9545 (c)

Intelligent Radio Signal Processing: A Survey

“…In the last decades, many conventional AMC methods have been proposed to enable dynamic spectrum access and intelligent spectrum management, where the expectationmaximization (EM) algorithms were employed to build maximum in likelihood-based classifiers [67]- [73]. Zhang et al [67] took advantage of the EM algorithm to estimate the maximum-likelihood of the unknown for modulation classification in a cooperative multiuser scenario.…”

“…The learning efficiency can be increased by recovering the centroids of all clusters by a constellationstructure-based reconstruction algorithm for parameter reduction with good convergence performance. Adaptive CSI estimation for modulation classification in MIMO systems was introduced by Abdul Salam et al [73] to jointly exploit the Kalman filter (KF) and an adaptive interacting multiple model (IMM). The IMM-KF output was subsequently analyzed by a quasi-likelihood ratio test (QLRT) algorithm for modulation identification.…”

“…Utilizing the same RadioML dataset, Huynh et al [5] applied their DL models to all the modulation types, whereas other researchers only considered a set of selected modulation formats, e.g., 5 types (BPSK, QPSK, 8PSK, 16QAM, 64QAM) [96], and 11 types (the RadioML2016.10a dataset) [188]. Additionally, other researchers tested their classification approaches using simulated datasets, wherein emulating realistic channel impairments is difficult [73], [79]. The fact is that, the higher the modulation order, the lower the classification accuracy [5].…”

AMC2-Pyramid: Intelligent Pyramidal Feature Engineering and Multi-Distance Decision Making for Automatic Multi-Carrier Modulation Classification

“…A. AMC without using DL The Kalman filter is integrated with adaptive interacting multiple models (IMM) [20] to estimate the channel state information (CSI). The channel is decomposed using singularvector decomposition (SVD) by adding up square root singular values.…”

“…For that, a classifier should be designed to classify a vast number of modulations. In earlier works, the following four steps are frequently used for multi-carrier signal modulation classification: [20] Achieves target optimal performance High complexity due to presence of noise Less classification accuracy due to lack of pre-processing PCA [21] Better performance during feature selection Less accuracy due to lack of significant features and pre-processing KNN [22] Low SNR rate for all modulations Less accuracy due to consider only instantaneous features M-QAM model [23] Reliable recognition rate even at low SNR rate (SNR<0) Only suitable for remove AWGN noise not suitable for other noise WF-RFT [24] Realistic recognition and high accuracy Only limited to PSK, ASK and QAM method when lower order cumulant is not suitable DAM [25] Rapid performance during modulation recognition for varying noise levels and modulations Not suitable for handling real time noisy data due to static threshold during classification Cepstrum model [26] Less computational complexity Inefficient feature extraction DSSS [27] Rapid performance during Tag recognition High information loss due to not select the number of PCA AMC with using deep learning CNN [28] Improved performance at low SNR rate (-4dB) Low performance in feature extraction CNN [29] High performance for modulation recognition with medium for high SNR rate Less accuracy due to lack of pre-processing CNN [30] High accuracy for digital signal modulation at low SNR (4dB) Not suitable for large datasets because SVM takes high amount of time for selecting modulation type CNN [31] Effective feature extraction Less accuracy at low SNR rate CNN [32] Reduce model size and accelerate computation Not suitable for all types modulation such as FSK, ASK, DPSK and etc.…”