An Unsupervised Neural Network Equalizer for Real-Time Photonics-Assisted Millimeter-Wave ISAC Systems With Error Propagation Suppression
Abstract
Integrated sensing and communication (ISAC) has emerged as a key paradigm for future wireless systems, where orthogonal frequency-division multiplexing (OFDM)/discrete multitone (DMT) is widely adopted in communication-oriented architectures. In the bidirectional (BiDi) ISAC system, the communication link is affected by nonlinear distortion and frequency-selective fading, while the sensing link adopts the two-way time-of-flight (TW-ToF) technique, whose ranging accuracy highly depends on stable signal reception and timestamp extraction. To enhance performance with minimal architectural modification, the compensation and equalization techniques at the receiver side have attracted considerable attention owing to their strong generality and ease of deployment. However, conventional decision-directed (DD)-least mean squares (LMS) equalizers are prone to error propagation, conventional blind equalization schemes often suffer from convergence and stability limitations, and supervised neural network-based equalizers usually introduce high computational complexity and implementation overhead. To address these issues, this article proposes a low-complexity blind deep neural network (DNN) equalization method applicable to both the communication and sensing links. The proposed equalizer is developed based on the DD criterion and incorporates a novel piecewise linear activation function, thereby enabling stable blind equalization. Compared with the conventional DD-LMS, the proposed method effectively suppresses error propagation. Furthermore, a real-time (RT) photonics-assisted millimeter-wave BiDi ISAC system based on field-programmable gate array (FPGA) is implemented and experimentally validated. Over a 236-m free-space transmission link, a DMT-16QAM communication data rate of 12.07 Gb/s is experimentally achieved, with a calculated theoretical range resolution of 0.508 cm. By cascading the proposed blind neural network equalizer with the conventional digital signal processing (DSP) chain, the bit error rate (BER) of the communication link can be reduced by up to 0.012, and the BER of the DMT-quadrature phaseshift keying (QPSK) signal in the sensing link remains error-free. This work explores high-performance RT ISAC receiver technologies and provides a valuable reference for the collaborative design of future high-capacity and low-latency ISAC systems.