Quantum-Inspired Machine Learning: A New Paradigm for Noise Filtering in Communication Systems

This project explores cutting-edge technologies integrating quantum computing and classical machine learning. It enhances error correction capabilities of wireless communication signals and reduces bit error rates through a hybrid quantum-classical architecture. The core idea is to use quantum-inspired algorithms for feature extraction and combine them with classical neural networks for decision-making, providing a new solution for noise filtering of QPSK-modulated signals.

In the evolution from 5G to 6G, traditional filtering techniques (Wiener, Kalman) perform poorly in complex scenarios such as non-Gaussian noise and impulse interference. Although deep learning can handle nonlinear mappings, it faces bottlenecks in high-dimensional quantum state feature representation. Quantum computing's superposition and entanglement properties can provide an exponential state space. Quantum-inspired algorithms (VQE, QAOA) draw on quantum frameworks but run on classical computers, retaining advantages while avoiding hardware limitations.

Architecture Design: A layered hybrid architecture consisting of a quantum-inspired feature extraction layer and a classical neural network layer, with the division of labor: "quantum for features, classical for decision-making". Quantum Feature Extraction: Uses a Variational Quantum Circuit (VQC) to encode QPSK constellation features into quantum states, extracting high-order statistical information via rotation gates/entanglement gates. Classical Network Integration: An attention-enhanced fully connected network, with a loss function combining cross-entropy and a bit error rate proxy term. Technical Implementation: Relies on tools like Qiskit/PennyLane, PyTorch/TensorFlow; preprocessing includes matched filtering, normalization, and quantum encoding; training uses an alternating optimization strategy (quantum gradients calculated via the parameter shift rule).

Experimental Setup: Test scenarios include mixed channels of AWGN, Rayleigh fading, and impulse noise; benchmarks are compared against traditional filtering, pure classical deep learning (CNN/LSTM), and pure quantum classifiers. Key Results: The hybrid method reduces BER by 15-30% in all scenarios, with more significant advantages under impulse noise; it has lower computational overhead than pure quantum methods. Visualization: The processed QPSK constellation point cloud is significantly tightened and returns to standard positions.

The project demonstrates the power of interdisciplinary integration of quantum physics, machine learning, and communication engineering. The hybrid architecture retains the advantages of quantum representation while avoiding quantum hardware limitations. For researchers, it is an entry point to understand quantum machine learning applications; for engineers, it provides a reusable framework; for decision-makers, it suggests that quantum advantages may be embodied in hybrid forms in specialized fields.

Potential Applications: Millimeter wave compensation for 5G/6G, massive MIMO channel estimation, NOMA multi-user detection, RIS-assisted communication optimization. Collaborative Evolution: As quantum hardware matures, it can be migrated to real quantum processors and combined with quantum communication technologies to enhance security and performance. Open Issues: Reducing quantum circuit depth to adapt to NISQ devices, designing efficient quantum-classical interfaces, and improving the interpretability of quantum feature extraction.