In this paper, we present a magnetic measurement system for integration into smart knee prostheses to accurately measure the combination of two knee rotations; namely flexion-extension (FE) and internal-external (IE) rotations. This measurement system consists of two permanent magnets inserted into the femoral and tibial parts of the prosthesis and a configuration of anisotropic magneto-resistive sensors placed in its polyethylene part. The sensor configuration was designed according to the sensitivity analysis. Several angle estimators were defined to obtain accurate angle estimations. These estimators ranged from different linear regression models to artificial neural networks. The estimators were trained and tested on several dynamic measurements of rotations of the prosthesis parts in a mechanical knee simulator also monitored using a stereophotogrammetry motion capture system. Considering the best estimators, the errors (mean ± SD) were 0.0° ± 0.9° and 0.2° ± 1.1° for IE and FE angle estimations, respectively. The imposed abduction-adduction (AA) rotations effect was investigated on the estimators in two cases: when the estimators were trained on data without AA (Case 1) and with AA (Case 2). The internal-external angle estimators showed high robustness to the imposed AA. The recorded errors for the best flexion-extension estimator were 0.9° ± 2.7° for Case 1 and 0.3° ± 1.7° for Case 2. The proposed system has thus demonstrated its ability to accurately estimate concurrent flexion-extension and internal-external rotations.