Abstract: Affected by non-ideal factors such as inverter nonlinearity and magnetic field space harmonics, the sliding mode observer(SMO) constructed based on the extended electromotive force(EEMF) model has problems such as high-frequency chattering and high-order harmonics in its output, which seriously affects the accuracy of rotor position estimation. A frequency adaptive filter(FAF) with band-pass frequency selection characteristics is proposed, and a cross-feedback network(CFN) is constructed to realize the back electromotive force(back-EMF) filtering. The filtering network has the characteristics of zero amplitude attenuation and zero phase delay to extract the fundamental component in the back-EMF observation value, and can eliminate the 5th and 7th harmonic components in the back-EMF observation value. The filtered fundamental back-EMF observation value is used to normalized phase-locked loop(NPLL) position estimation to reduce the position pulsation error and achieve high-precision rotor position estimation. Finally, the effectiveness of the proposed method is verified by a 1.5 kW sensorless control experimental platform of interior permanent magnet synchronous motor(IPMSM).