3 January 2024 in IEEE Transactions on Aerospace and Electronic Systems
Waveform design with constant modulus constraint (CMC) is of great importance in multiple-input multiple-output (MIMO) radar systems. Both the relaxations in model-based waveform design methods and data limitation in data driven Deep Neural Networks (DNNs) methods could result in performance degradation. Nevertheless, these two types of methods have their unique advantages. This motivates us to develop a model-based Learned Complex Circle Manifold Network (LCCM-Net) method without relaxation, by exploiting the advantages of both gradient descent models over CCM and DNNs with limited data for learning. More concretely, we propose to formulate the waveform design problem as an Unconstrained Quadratic Fractional Problem (UQFP) on the CCM. To solve the resultant problem, the gradient descent algorithm is unfolded as the network layer over the CCM, and the step sizes are adaptively learned. Furthermore, for discrete phases adopted in practical systems, we develop a Quantized LCCM-Net (QLCCM-Net), where a low-resolution nonuniform quantizer is designed to quantize the phase. This quantizer relies on a uniquely designed soft staircase function, which incorporates learnable parameters that allow it to adaptively fine-tune the decision region. The performance superiority of the proposed method is evidenced by comparing with the existing methods.
22 February 2024 in IEEE Geoscience and Remote Sensing Letters
Multiple Input Multiple Output (MIMO) radar unimodular waveform design with range-ISL optimization is a key technology in remote sensing. Due to the non-convex quartic objective function and constant modulus constraint (CMC), the problem is NP-hard and non-convex. Existing methods mainly include relaxation methods or non-relaxation methods with huge computational cost. We notice that complex circle manifold (CCM) naturally satisfies the CMC. By projecting onto the CCM, the problem is transformed into an unconstrained minimization problem that can be addressed using the Riemannian gradient descent (RGD) algorithm. Furthermore, we notice that the RGD algorithm can be unfolded into a deep learning model. Hence, a computationally efficient method without relaxation, Iterative Deep Unfolding Network (IDUN), is proposed. First, this problem is converted into an unconstrained fourth-order polynomial minimization problem on the CCM. Then, by unfolding RGD algorithm as the network layer, IDUN is developed with adaptively learning the step sizes. Compared with existing methods, the proposed method has superior performance and less computational cost.