Multichannel noise cancellation uses witness channels to predict noise in a primary measurement channel. Target-signal leakage into the witnesses can bias transfer-function estimates and distort the recov-ered signal. We propose mask-constrained sequential Wiener projection (MC-SWP), which combinesfrequency-band protection with frequency-dependent sequential updates. Transfer estimates around thesignal frequencies are replaced by complex interpolation from both sides of each protected band, retain-ing nonzero in-band noise prediction. The method selects witness features sequentially and updates theprimary residual and remaining features using frequency-dependent coefficients. We evaluate MC-SWPusing controlled injections into 22 public structural-vibration records and two measured datasets from thesame precision-measurement apparatus. A mask-on/off ablation within MC-SWP shows reduced leakage sensitivity and signal distortion while retaining noise suppression near the signal frequencies. Ata leakage amplitude ratio of 1, a separate comparison with identical spectral-averaging and protected-band widths gives a median band-of-interest residual RMS amplitude of 4.88% of the raw amplitude forMC-SWP, compared with 9.24% for protected polynomial ridge regression. The corresponding incremen-tal signal-response error increases from 3.19% to 4.74%. Both measured datasets also show lower noiseresiduals with MC-SWP than with the comparison methods. These results support frequency-band protec-tion and sequential projection for noise cancellation in the presence of signal leakage, while quantifyingthe accompanying signal-distortion trade-off.

Noise cancellation,Wiener filtering,Witness channels,Signal leakage,Frequency-band protection,Sequential projection