A Signal-Adaptive Semi-Supervised Framework for Robust ISAC Under Hardware Impairments

Integrated Sensing and Communications (ISAC) is a technology for next-generation wireless systems. However, its practical deployment is fundamentally limited by the time-varying nature of hardware impairments, which renders conventional static compensation methods ineffective and degrades both sensing and communication quality. To address this limitation, we propose a signal-adaptive semi-supervised learning (SSL) framework to robustly compensate for dynamic hardware impairments. The core of our approach is a learnable architecture, PertNet, which incorporates an LSTM-based Temporal Encoder to extract temporal features from each signal snapshot and dynamically adjust hardware perturbation parameters. By using PertNet to continuously learn and adjust for signal distortions in real-time, our method solves the critical problem of dynamic, time-varying hardware impairments that conventional static compensation techniques cannot address. Experimental results demonstrate that our framework reduces the error from 1.2 m to 0.5 m compared to a non-adaptive baseline.