Learning-Enhanced Riemannian Gradient Descent Method for Transmit-Receive Joint Design Towards ISRJ Suppression

The interrupted sampling repeater jamming (ISRJ) can create false targets that obscure real targets, leading to radar target detection failures. This study investigates the ISRJ countermeasure in multiple-input multiple-output (MIMO) radar through transmit-receive joint design. We model the transmit-receive design problem as a jointly constrained optimization problem, aiming to minimize the waveform sidelobes, ISRJ energy, and mutual interference among various waveform-filter pairs. To address the difficulties posed by non-convex constraints, we transform the original constrained problem in Euclidean space into an unconstrained one in Riemannian manifold space. To simultaneously and adaptively update the transmit waveforms and receive filters, we propose a learning-enhanced Riemannian gradient descent (LE-RGD) method, which unfolds the classical Riemannian gradient descent (RGD) method into layers of a neural network. The LE-RGD algorithm directly optimizes transmit waveforms and receive filters through implicit gradient descent iterations, where the optimization strategy is dynamically and adaptively determined by a parameterized network at each iteration. Furthermore, the LE-RGD network is randomly initialized at each problem instance and updated iteratively, facilitating its application in diverse jamming environments without the need for labeled training data. Numerical experiments conclusively show that the LE-RGD method can effectively design transmit waveforms and receive filters with high performance in terms of pulse compression and ISRJ suppression.