Machine Learning for Integrated Sensing and Communications under Modeling Mismatch

Doktorsavhandling, 2026

Integrated sensing and communication (ISAC) is envisioned as a key functionality of sixth-generation wireless systems, enabling spatial perception of the environment through the same hardware, spectrum, and waveforms used for data transmission. Realizing this vision is challenged by modeling mismatch between assumed signal models and real operating conditions, which conventional model-based signal processing struggles to address. This thesis studies two such forms of mismatch: (i) hardware impairments and (ii) complex propagation conditions such as non-line-of-sight. This thesis investigates machine learning as a data-driven complement to model-based signal processing in ISAC, with contributions spanning supervised, semi-supervised, unsupervised, and self-supervised learning regimes across four appended papers.

Paper A proposes a model-based end-to-end learning framework that jointly optimizes the ISAC transmitter and sensing receiver under antenna-array hardware impairments, enabling supervised calibration through differentiable ISAC algorithms. The results show that learning the parameterized impairments outperforms standard model-based calibration while generalizing to unseen scenarios. Paper B reduces the labeling cost of this framework, showing that semi-supervised learning matches fully supervised performance with far less labeled data, while purely unsupervised learning falls short. Paper C closes this gap in the multi-target case, calibrating transmitter and receiver impairments with no labeled data via an approximation of the channel gradient, performing close to supervised calibration and generalizing better across signal-to-noise ratios.

Paper D removes the reliance on parametric line-of-sight channel models, addressing active positioning through a self-supervised channel-charting (CC) framework augmented with a digital twin. Matching large-scale channel state information features, the proposed method outperforms the CC state of the art and remains robust to modeling mismatch and distribution shifts.