This paper proposes a strategy using new transformation matrices to calculate new current references when a non-sinusoidal seven-phase permanent magnet synchronous machine (PMSM) has an open-circuited phase. The new current references allow to obtain a smooth torque with similar copper losses in the remaining healthy phases even when the back electromotive force (back-EMF) is non-sinusoidal. A real-time current learning process using an adaptive linear neural network (Adaline) is applied to extract from measured currents useful harmonic components in torque generation. It improves torque quality, especially at high speed, even when standard proportional-integral (PI) controllers are applied. In addition, similar copper losses in the remaining phases with the new current references can avoid overheating of windings. The effectiveness of the proposed control strategy is validated by numerical results.