Automatic Modulation Classifier Using Artificial Neural Network Trained by PSO Algorithm

Li Cheng, Jin LiuJournal of Communications 2013 · DOI

TL;DR

This paper tackles the challenge of reliably identifying different types of digital communication signals, a crucial step for efficient signal demodulation. The authors propose a novel approach by combining the strengths of Artificial Neural Networks (ANNs) with Particle Swarm Optimization (PSO) to train the ANN, overcoming the limitations of traditional training methods and achieving high recognition accuracy even in noisy environments.

What Problem Does This Paper Address?

In modern wireless communication, signals are transmitted using various modulation schemes (e.g., ASK, PSK, FSK). To correctly receive and interpret these signals, a system needs to first identify which modulation scheme was used. This process is called Automatic Modulation Classification (AMC). Traditional methods for AMC often rely on complex mathematical models or feature extraction followed by classification. Artificial Neural Networks (ANNs), particularly the Backpropagation Neural Network (BPNN), have shown promise for this task due to their ability to learn complex patterns. However, BPNNs suffer from two major drawbacks: they can easily get stuck in "local minima" during training, meaning they don't find the best possible solution, and their performance is highly dependent on the initial random settings of their internal weights. This sensitivity makes them unreliable and can lead to poor recognition accuracy, especially in low signal-to-noise ratio (SNR) conditions.

Methods

The core of this paper's contribution lies in its hybrid approach: PSO-BPNN.

  • Backpropagation Neural Network (BPNN): This is the chosen classifier architecture. BPNNs are feed-forward networks that learn by adjusting their internal weights based on the error between their predicted output and the actual target output. They are known for their ability to approximate complex functions.
  • Particle Swarm Optimization (PSO): This is a metaheuristic optimization algorithm inspired by the social behavior of bird flocking or fish schooling. PSO is used to train the BPNN. Instead of relying on gradient descent (which can get stuck in local minima), PSO explores the solution space more broadly to find a global optimum. In this context, PSO is used to find the optimal weights and thresholds for the BPNN, aiming to minimize the classification error.
  • Feature Extraction: Before feeding signals into the classifier, specific features are extracted. The paper utilizes six features derived from the instantaneous amplitude, phase, and frequency of the received signal. These features are designed to help distinguish between different modulation types.

The overall process involves pre-processing the received signal, extracting these six key features, and then using the PSO-trained BPNN to classify the signal into one of the seven supported modulation types.

Datasets

The paper does not explicitly mention using pre-existing, publicly available datasets for training and evaluation. Instead, it describes generating simulated signals for seven digital modulation types (2ASK, 4ASK, BPSK, QPSK, 2FSK, 4FSK, OFDM).

  • Training Data: 600 samples were generated for each modulation scheme. 400 samples were used for training the PSO-BPNN.
  • Testing Data: The remaining 200 samples for each scheme were used for testing.
  • Noise Environment: The simulations were conducted in an Additive White Gaussian Noise (AWGN) environment.
  • SNR Range: Experiments were performed across a range of Signal-to-Noise Ratios (SNRs) from 0 dB to 20 dB, with 5 dB intervals.

Key Insights and Ideas

  • PSO as a Global Optimizer for BPNN: The paper convincingly demonstrates that PSO can effectively overcome the local minima problem inherent in BPNN training. By using PSO to find optimal weights, the network's ability to generalize and perform well across different conditions is significantly enhanced.
  • Robustness in Low SNR: A critical finding is the classifier's strong performance even at very low SNRs (e.g., 0 dB). This is a significant advantage for real-world applications where signal quality can degrade.
  • Feature Engineering for Classification: The selection of six specific features derived from signal characteristics (amplitude, phase, frequency) is shown to be effective for distinguishing between the seven modulation types.
  • Trade-off between Iterations and Convergence: While PSO-BPNN achieves better recognition rates, the paper also notes that other methods like the Lenvenberg-Marquardt (L-M) algorithm might converge faster in terms of iterations when the network is smaller, but PSO-BPNN offers superior performance and stability, especially at lower SNRs.

Key Results

  • High Recognition Accuracy: The proposed PSO-BPNN classifier achieved approximately 92% recognition performance at an SNR of 0 dB.
  • Improved Performance with SNR: As the SNR increased, the recognition rate further improved, reaching close to 100% at higher SNRs (e.g., 15 dB and 20 dB).
  • Comparison with other methods: The PSO-BP classifier consistently outperformed standard BP classifiers and the L-M classifier in terms of recognition rate, particularly at lower SNRs.
  • Efficiency: The paper claims the PSO-BP classifier can reduce calculation time and is simple to implement.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that some modulation types, like 2ASK and 4ASK, can still be confused due to their similar feature characteristics (specifically, the standard deviation of the normalized-centered instantaneous amplitude, σaa). Similarly, QPSK can be misidentified as BPSK or 2FSK.
  • The determination of the optimal number of hidden layer neurons in the BPNN was done through simulation and trial-and-error, suggesting a need for more automated methods.
  • Future Work:
    • The authors suggest exploring other bio-inspired theories (beyond PSO) to further improve BPNN training for AMC.
  • They also propose investigating dynamic adjustment of the number of nodes in the BPNN's hidden layer.

Summary updated

Abstract

The back propagation (BP) neural networks have been commonly used for automatic modulation recognition since the late 1990s. However, the back propagation algorithm easily falls into local minimum and the network learning is sensitive to initial weight values which usually determined by experience. The particle swarm optimization (PSO) algorithm is a global heuristic searching technology. By combing these two techniques, this paper presents a novel automatic modulation classifier. By training BP neural network with PSO optimization, this classifier can identify seven digital signals with six input features. The experiment results show that this recognizer achieves about 92% recognition performance at the SNR of 0dB.
BibTeX
@article{cheng2013automatic,
  author = {Li Cheng and Jin Liu},
  title = {{Automatic Modulation Classifier Using Artificial Neural Network Trained by PSO Algorithm}},
  journal = {Journal of Communications},
  year = {2013},
  doi = {10.12720/jcm.8.5.322-329},
}
Figure